A causality-enhanced multiresolution residual learning framework for image retrieval with fast osprey optimization.

A causality-enhanced multiresolution residual learning framework for image retrieval with fast osprey optimization - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 13;16:17258. doi: 10.1038/s41598-026-48612-1 A causality-enhanced multiresolution residual learning framework for image retrieval with fast osprey optimization Abdulrahman Yousif Zeain 1, ✉ , Abdullahi Abdu Ibrahim 1 Author information Article notes Copyright and License information 1 Institute of Graduate Studies, Electrical and Computer Engineering, Altınbaş University, Istanbul, Turkey ✉ Corresponding author. Received 2025 Jul 2; Accepted 2026 Apr 8; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13234330  PMID: 41974818 Abstract The rapid expansion of large-scale image collections demands content-based image retrieval systems that are both semantically reliable and computationally efficient. Conventional CBIR approaches based on isolated visual descriptors often fail under high intra-class variation and inter-class ambiguity. This work proposes a causality-enhanced multiresolution residual learning framework that integrates deep representation learning, causal modeling, and lightweight optimization for robust image retrieval. A multiscale ResNet 50 backbone captures complementary fine-grained and high-level semantic features across multiple resolutions, while a causal variational autoencoder explicitly models latent spatial dependencies to improve semantic consistency and interpretability. Feature redundancy is minimized by an enhanced fast osprey optimization algorithm, enabling compact, discriminative feature selection at low computational cost. Extensive experiments on CIFAR-10, Oxford Flowers, and Corel 1000, using fivefold cross-validation, demonstrate consistent improvements over recently published thirteen state-of-the-art methods. Further gains in mean average precision, normalized discounted cumulative gain, and reduced retrieval time confirm the effectiveness, robustness, and practical viability of our approach across diverse retrieval scenarios. Moreover, the retrieval samples exhibit strong robustness against high intra-class variability and severe inter-class ambiguity, preserving discriminative consistency across visually similar categories. Keywords: Content-based image retrieval, Causality, Osprey optimization algorithm, RESNET-50, Variational autoencoder Subject terms: Engineering, Mathematics and computing Introduction Content-based image retrieval is an advanced computer vision technique that automatically retrieves images from databases based on their visual content, rather than relying on textual metadata or keywords 1 . This technique analyses the attributes of color, texture, shape, and spatial relationships in images to form a digital signature, or fingerprint, for each image 2 . It enables efficient, accurate matching and retrieval of images that semantically correspond to the query image in the database. The classic systems of image retrieval, primarily based on metadata or textual annotations, face numerous difficulties in identifying similar images due to the increased complexity and scalability of contemporary image datasets. Images carry vast amounts of information beyond simple descriptions, so systems are inherently needed to accurately capture and retrieve visual content based on the rich, multidimensional features of images. Deep learning-based CBIR frameworks have attracted attention for their potential to revolutionize image retrieval by automatically extracting, learning, and utilizing high-level semantic features. However, current deep learning models frequently encounter scalability, speed, and precision issues on large-scale image datasets 3 . Deep learning architectures, especially convolution neural networks (CNNs), have rapidly become the core of modern CBIR systems because of their exceptional capabilities to learn rich, complex features from images automatically 4 , 5 . Unlike design, where traditional methods differ based on predefined feature-extraction algorithms, CNNs can discover a range of features, from bare edges and textures to abstractions such as objects, scenes, and emotions within images. This hierarchical feature extraction has enhanced the Precision and effectiveness of CBIR systems by enabling them to comprehend images at various levels of abstraction 6 . Existing deep learning based CBIR approaches can be broadly classified into three categories based on their representation learning strategy. The first category includes discriminative CNN-based models, where features extracted from architectures such as VGG, Inception, and ResNet are directly used for similarity-based retrieval 7 – 9 . The second category comprises metric-learning and hashing-based deep networks that learn compact embedding spaces optimized for retrieval via contrastive or triplet losses 10 , 11 . These models focus on maximizing inter-class separation and minimizing intra-class variation, which directly benefits retrieval accuracy. The third category comprises generative and representation learning approaches 12 , such as autoencoders and variational autoencoders, which learn compact latent codes that preserve semantic structure while reducing redundancy and noise. Despite these advances, deep learning-based CBIR systems still face significant challenges in scalability, computational cost, and robustness. Training and deploying deep networks on large-scale datasets is computationally expensive and memory-intensive 13 . Moreover, deep features are often sensitive to noise, illumination changes, and variations in input data, which can degrade retrieval reliability in practical environments 14 . In addition, most CNNs operate at a fixed spatial resolution, which limits their ability to capture both fine textures and large-scale structures simultaneously. To address these issues, multiresolution CNN architectures, such as ResNet-based frameworks, have been introduced to extract features at multiple spatial scales, thereby capturing both local and global information 15 . Although this multiresolution modeling improves retrieval performance across diverse image categories, it also increases feature dimensionality and redundancy, which can negatively affect computational efficiency and retrieval speed 16 . It increases computational complexity and may slow retrieval times, which is especially undesirable in real-time applications. Therefore, a newly developed Causal Variational Autoencoder (VAE) was integrated with the ResNet-50 architecture to ensure that the latent representations explicitly encode causal relationships between variables. The primary reason for selecting ResNet-50 in our proposed work is its ability to efficiently extract high-level semantic features while maintaining a balance between depth and computational efficiency. Moreover, integrating ResNet-50 into our fast osprey optimization algorithm 17 (FOOA) enhances similarity matching by leveraging its deep feature representations to improve retrieval accuracy. Hence, ResNet-50 gives the best tradeoff between computational efficiency and retrieval performance. The proposed work integrates multiresolution analysis and optimization strategies to develop a more accurate and effective CBIR system. Therefore, the system can capture all details at different resolutions by integrating multiscale feature extraction, enabling highly meaningful and precise semantic retrievals. Optimization algorithms further ensure that the system remains computationally efficient and scalable regardless of dataset size. These advances represent a significant leap forward in reducing the deficit of traditional deep learning-based CBIR systems, making them highly robust and scalable for current image retrieval tasks. The proposed method is tested on the CIFAR-10 18 , Oxford Flower 19 , and Corel-1000 datasets 20 using fivefold cross-validation and two similarity metrics: structural similarity index (SSIM) and histogram intersection (HI). In addition, various performance metrics, such as accuracy, precision, recall, and F1-score, are used to evaluate the model’s effectiveness and compare its results with those of state-of-the-art models and algorithms, with superiority reported. Literature background Content-based image retrieval, a prominent emerging technology in managing and retrieving visual information from huge image repositories over the last decade, has gained significant importance. The efficiency and Accuracy of retrieval systems become increasingly important as the size of image databases increases exponentially, across domains such as medical imaging and e-commerce applications 21 , 22 . These innovations, particularly convolutional neural networks, have signature effects that reduce the semantic gap between low-level visual features and high-level semantic concepts in CBIR systems. A concise study of several recently developed CBIR frameworks is summarized in Table 1 . Table 1. Tabular representation of recently published state-of-the-art research studies. Authors Proposed methodology Results Gaps/limitations Luo et al. (2022) 23 CNN with ResNet-50 + attention module; Locality Sensitive Hashing (LSH) for indexing Outperformed prior methods on Corel-1K and Corel-5K No mention of semantic gap reduction or complex datasets Christy and Dhanalakshmi (2022) 24 Geon similarity + Grey Wolf Optimization (GWO) + novelty-driven ResNet-50-based CNN for hashing and classification Achieved high precision, recall, accuracy, and mAP High model complexity due to multiple hybrid integrations Bibi et al. (2022) 25 VGG-19 is used for deep feature extraction, GA is used for feature selection, and ELM is used for classification Superior performance on five datasets; validated via the Wilcoxon test Lacks mention of real-time retrieval speed or storage efficiency Rastegar et al. (2024) 26 RBF neural network for EEG classification optimized by Jellyfish Search for Gaussian center selection, combined with Locally Linear Embedding for dimensionality reduction Achieved consistent performance improvement over classical RBFNN and CNN-based EEG classifiers across challenging EEG datasets Local-only feature-based methods fail with diverse image datasets Rastegar and Giveki (2023) 27 Deep CNN-based visual feature extraction integrated with relevance feedback and generalized discriminant analysis for refined class separation Achieved mAP of 96.8% on OT dataset, 95.62% on Corel one thousand, and 98.78% on Caltech 101, demonstrating strong retrieval and recognition accuracy High complexity Giveki et al. (2024) 28 Novel boundary-aware discriminative shape descriptor integrated with enhanced bag of features using a new codebook generation strategy Demonstrated superior robustness to both inter-class and intra-class shape variations compared to existing shape descriptors It may not handle noisy or ambiguous queries effectively Giveki and Esfandyari (2024) 29 Unified image representation by fusing InceptionResNet V2 semantic features with complementary color and texture descriptors Achieved consistent superiority on both single-label and multi-label benchmarks for recognition and retrieval tasks Focused only on model comparison; lacks novel optimization Giveki et al. (2024) 30 Antlion optimization-driven automatic CNN architecture and hyperparameter design for plant leaf disease recognition Delivered consistent performance gains over state-of-the-art deep learning models across four agricultural datasets Potentially high feature dimensionality; complexity in fusion Farsijani et al. (2025) 31 Enhanced U Net with ResNet-50 encoder and channel spatial attention in the decoder for preserving edge and texture information Achieved IoU of 93.35% and Dice score of 96.45%, confirming high precision plant disease segmentation High computational cost due to a complex pipeline Singh and Singh (2025) 32 ResNet-18 + ResNet-50 feature fusion (ResNetFuse) + t-SNE for dimensionality reduction 96.15% mAP; 87.5% reduction in feature size Fusion increases initial complexity and requires dimensionality reduction Rani et al. (2025) 33 Separable CNN with intermediate feature extraction + ANNOY indexing Near-perfect recall and accuracy Lacks evaluation across diverse datasets or scalability analysis Open in a new tab Luo et al. (2022) 23 proposed an enhanced CNN-based image search method that incorporates attention and locality-sensitive hashing. ResNet-50 is used as the feature extractor, with an attention module appended after the convolutional layers to strengthen feature attention. The fully connected layer features are subsequently hashed using locality-sensitive hashing for dimension reduction and to form an efficient index. Similarity between the query and database features is utilized to perform the retrieval. The approach was tested on Corel-1K and Corel-5K datasets, outperforming competitive methods. Christy and Dhanalakshmi (2022) 24 utilized the Geon similarity model for efficient and precise calculation of image similarity. A novel Grey Wolf Optimization (GWO) method and a novelty-driven ResNet-50-based CNN were combined for hashing and classification. The method achieved better recognition accuracy and precision than state-of-the-art techniques, providing robust performance across precision, recall, accuracy, and mAP metrics. Bibi et al. (2022) 25 proposed a CBIR approach that utilizes VGG-19 for deep feature extraction, GA for feature selection, and ELM for rapid classification. The method extracts both local and global image information, minimizes computational burden, and outperforms state-of-the-art techniques on five datasets. Wilcoxon test results validate their statistical significance. Rastegar et al. (2024) 26 introduced an RBFNN-based EEG classifier in which Jellyfish Search optimizes the selection of Gaussian centers, and Locally Linear Embedding reduces signal dimensionality. Extensive experiments on challenging EEG datasets show consistent performance gains over state-of-the-art RBFNNs and even CNN-based approaches. Rastegar and Giveki (2023) 27 integrated a relevance feedback strategy with a deep CNN to extract visual features and discriminated them using generalized discriminant analysis. Validation results obtained from the OT, Corel-1000, and Caltech-101 confirm superior mean Average Precisions (mAPs) of 96.8%, 95.62% and 98.78%, respectively. Giveki et al. (2024) 28 proposed a discriminative shape descriptor that captures boundary-level variations and integrates it with an enhanced bag-of-features model using a novel codebook generation strategy. Experimental results on complex-shape datasets demonstrate superior robustness to inter- and intra-class variations compared to existing descriptors. Giveki and Esfandyari (2024) 29 introduced a unified image representation that fuses InceptionResNet V2-based semantic features with complementary color and texture cues. Extensive evaluations on multiple single-label and multi-label benchmarks confirm robust performance and consistent superiority over state-of-the-art recognition and retrieval methods. Giveki et al. (2024) 30 employed antlion optimization to automatically design CNN architectures and tune critical hyperparameters for the recognition of plant leaf diseases. Evaluations across four challenging agricultural datasets demonstrate consistent performance gains over state-of-the-art deep learning methods, highlighting the robustness of the approach to evolving disease patterns. Farsijani et al. (2025) 31 proposed an enhanced U-Net architecture using a ResNet-50 encoder and channel-wise spatial attention in the decoder to preserve fine-grained edge and texture information. The model achieves an IoU of 93.35% and a Dice score of 96.45%, demonstrating superior accuracy in segmenting plant leaf diseases. Singh and Singh (2025) 32 introduced ResNetFuse by combining features from ResNet-18 and ResNet-50 with t-SNE-based reduction, achieving a 96.15% mAP on the KVASIR dataset while reducing feature dimensionality by 87.5%, making it suitable for resource-constrained settings. Rani et al. (2025) 33 proposed a lightweight separable CNN with intermediate feature aggregation and ANNOY-based retrieval, achieving near-perfect recall and accuracy. However, existing CBIR methods still face challenges, such as semantic ambiguity, feature redundancy, and limited interpretability, due to high intra- and inter-class overlap. Motivation and research gaps The majority of existing work relies on single-scale representations or heuristic optimization, without a principled framework that integrates multiresolution learning, causal feature interpretation, and robust optimization. To fill these gaps, our framework introduces a multiscale ResNet-50 to extract both global and fine-grained semantic features, a Causal Variational Autoencoder (CausalVAE) to reveal disentangled, interpretable latent patterns, and an improved Fast Osprey Optimization Algorithm to eliminate redundancy and select the most discriminative features. This integrated design addresses the shortcomings identified in previous studies by providing a more robust, scalable, and semantic-aware solution for content-based image retrieval. Our contribution The main contributions of this work are summarized as follows: A multiresolution learning framework is integrated with the ResNet-50 architecture to effectively extract both fine-grained and high-level image representations at different spatial scales, improving retrieval accuracy in content-based image retrieval systems. A multiresolution-oriented ResNet-50 architecture is combined with a Variational Autoencoder to enhance the interpretability of deep representations and capture meaningful latent spatial dependencies within images. A fast variant of the Osprey Optimization Algorithm is developed to perform efficient feature selection and reduce redundant feature representations, thereby improving system scalability and computational efficiency. The integration of multiresolution feature learning with optimization significantly enhances the robustness of the retrieval framework when handling complex and non-uniform image distributions. The effectiveness of the proposed framework is validated on three benchmark datasets—CIFAR-10, Oxford Flower, and Corel-1000—using multiple evaluation metrics including Accuracy, Precision, Recall, and F1-score. Experimental results In this section, we present the results of experiments conducted to evaluate the performance of our proposed method. The performance of the proposed method is also compared with state-of-the-art methods to demonstrate its effectiveness for image retrieval. The detailed discussion is given in the subsequent subsections: Experimental setup The experimental evaluation of the proposed CBIR framework was conducted in a controlled, reproducible computing environment. All experiments were conducted on a Linux-based workstation running Ubuntu 20.04 LTS, equipped with NVIDIA GPUs, each with at least 12 GB of GPU memory, a multi-core CPU, 64 GB of system RAM, and high-speed SSD storage to support large-scale feature extraction and retrieval. The proposed deep learning models were implemented using TensorFlow version 2.8 with GPU acceleration. Data preprocessing and feature handling were performed using NumPy, Pandas, and Scikit learn. Model training was controlled using TensorFlow callbacks, including early stopping, adaptive learning rate scheduling, and model checkpointing, to ensure stable convergence and prevent overfitting. For retrieval evaluation, a fixed number of query images were randomly selected from each class to ensure balanced and unbiased testing. Specifically, 100 query images per class were used for CIFAR-10, resulting in a total of 1000 queries. For Oxford Flowers, 10 query images were used per category, resulting in a total of 1020 queries. Additionally, 10 queries per category were selected for Corel 1000, resulting in a total of 1000 queries. For each query, the entire dataset was searched, and retrieval performance was measured at different retrieval depths ( k ) of 10, 20, and 50. All reported precision, recall, and F-score values correspond to the standard top 20 setting. To ensure statistical reliability, all experiments were repeated five times using different random query selections, and the mean and standard deviation of all metrics were computed. Additionally, paired t tests 34 were performed to compare the proposed method with each baseline method and verify the statistical significance of the performance improvements. In all cases, the proposed framework achieved statistically significant gains, with p -values < 0.01. Dataset details We validate our proposed framework using three diverse datasets: CIFAR-10, Oxford Flowers, and Corel-1000. The CIFAR-10 dataset is a popular benchmark dataset for computer vision, comprising 60,000 32 × 32-pixel color images. The images are uniformly distributed in 10 classes, with 6000 images per class. The dataset is split into 50,000 training images and 10,000 test images. The 10 classes of CIFAR-10 are common objects and animals: airplanes, automobiles, birds, cats, deer, dogs, frogs, horses, ships, and trucks. These classes do not overlap, so no object is in more than one category. A few sample images from the CIFAR-10 dataset are shown in Fig. 1 . Fig. 1. Open in a new tab Sample CIFAR-10 dataset images 35 . The Oxford Flower dataset is another widely used benchmark for fine-grained image classification, especially within botanical image recognition. It contains 8189 images representing 102 flower classes. There are 40–258 images per class, making the dataset unbalanced. The flowers were selected based on their variety, visual resemblance, and presence in the United Kingdom. A few sample images from the Oxford Flower dataset are shown in Fig. 2 . The Corel-1000 dataset contains 1000 images across 10 categories: beaches, buses, dinosaurs, flowers, and horses. There are 100 images per category, making it ideal for image classification and content-based image retrieval. Images have varying resolutions, usually about 256 × 384 or 384 × 256 pixels. It contains a variety of objects and scenes, with intra-class variations and background clutter added. It enables benchmarking of texture, color, shape, and deep-feature-based approaches. It is utilized by researchers to test machine learning and deep learning models. A few sample images from the Corel-1000 dataset are shown in Fig. 3 . Fig. 2. Open in a new tab Sample Oxford Flower dataset images 36 . Fig. 3. Open in a new tab Sample Corel-1000 dataset images. Performance evaluation metrics For a proper understanding of the proposed CBIR model and meaningful comparison among different approaches, the following evaluation metrics must be specified along with benchmark datasets. Accuracy is the most widely used measure for comparing the performance of classification. It provides the proportion of correct classifications out of the total instances 37 . Precision : It is the proportion of truly positive instances against the number of cases predicted as positive. This means that the ratio of True Positive (TP) predictions to the model’s total number of positive predictions is the definition of accuracy 38 . It can be calculated using Eq. 1 . 1 where True Positive (TP) and False Positive (FP) refer to how many cases the model correctly classifies as positive and how many occurrences the model misclassifies as positive when they are negative. Recall-sensitivity or true positive rate : It is the ratio of truly positive instances to the sum of the true positive and false negative ratios 39 . It measures how well a classifier performs in detecting all positive instances. Mathematically, Recall is calculated as: 2 F1-score : It is the harmonic mean of Precision and Recall 40 . It occurs when there is an imbalanced class distribution problem or when one desires to balance precision and Recall. Mathematically, Recall is calculated as: 3 Mean Average Precision (mAP) : It measures the ranking quality of the returned results by averaging precision across several recall levels 41 . A higher mAP indicates better ranking and Higher Precision in retrieving effective images or objects. It can be calculated using Eqs. ( 4 ) and ( 5 ). 4 5 where Q refers to the total number of test queries used for evaluation. shows the precision averaged over different recall levels for a single query. The total number of relevant items for a query ( ) determines the number of results retrieved that are actually relevant for that particular query. The rank location ( ) refers to where an element occurs in the list of retrieved items. Precision at rank estimates the proportion of useful results found up to rank k , so high Precision at higher ranks. The difference in Recall at rank k (ΔR(k)) reflects how Recall increases when a useful item is in the retrieved list. Precision–Recall Curve : Precision–Recall curves are crucial metrics for evaluating classification model performance, particularly with imbalanced datasets 40 . They graph Precision versus Recall, showing the tradeoff between the two measures as the decision threshold changes. Generally, as Recall improves, Precision worsens, reflecting the difficulty of having high Accuracy while detecting more true positives. Normalized Discounted Cumulative Gain (NDCG) : It measures the performance of ranking algorithms in information retrieval. It is determined by dividing the Discounted Cumulative Gain 42 (DCG) by the Ideal Discounted Cumulative Gain (IDCG), yielding a score ranging from 0 to 1. The greater the NDCG value, the better the ranking performance, and it emphasizes both relevance and the ranking order of the retrieved items. In our work, two similarity measures, (1) Structural Similarity Index (SSIM) and (2) Histogram Intersection, are used to calculate the similarity between a query and testing samples. SSIM was selected because it is widely used to measure perceptual similarity between images. Unlike pixel-wise metrics such as mean squared error, SSIM jointly models luminance, contrast, and structural information, which are key components of human visual perception and are well aligned with multiresolution deep feature representations. Histogram Intersection was employed as a complementary metric because it effectively measures the distribution level similarity between feature histograms and is robust to variations in feature scale. Conventional distance metrics, such as Euclidean and cosine similarity, were also examined but proved less effective at capturing fine-grained perceptual and distributional differences in the learned feature space. It can be mathematically defined as: 6 Here, and are the mean values of the image and , respectively, while are the variances of and . represents the covariance between and , and ​ and ​ are small constants added to stabilize the division when the denominators are close to zero. SSIM is calculated over small image windows; thus, it captures local structural information and provides a much more perceptually accurate similarity measure. The second method, histogram intersection, is a straightforward and efficient image similarity metric that calculates the overlap between the histograms of two images. It calculates the sum of minimum values between corresponding histogram bins, thereby enabling one to find the similarity between two histogram plots. This method is more straightforward in comparing color histograms or grayscale images, as it does not depend on absolute pixel values but rather on the distribution of pixel intensities across the image. It can be mathematically defined as: 7 where and are the histogram values (or frequencies) for image one and image two at bin , respectively. The sum is taken over all histogram bins . The value of the Histogram Intersection ranges from 0 to 1, where 0 indicates no overlap between the histograms and 1 indicates complete overlap, meaning the images are identical in their pixel distributions. Results on the CIFAR-10 dataset For each of the ten classes, the performance of the proposed CBIR model on the CIFAR-10 dataset is analyzed, revealing class-specific retrieval efficiency (Table 2 ). The similarity measures SSIM and Histogram Intersection were used to compare performance across all classes, with Accuracy, Precision, Recall, and F-score values ranging from 0.95 to 1.00. Our model achieves 0.979 accuracy on the Airplane class with SSIM and 0.975 with Histogram Intersection. Based on precision and recall values, the model can efficiently differentiate between airplanes and other objects by leveraging unique shapes and background patterns specific to this class. Similarly, the Accuracy for the Automobile class reaches 0.982 (SSIM) and 0.978 (Histogram) when wheels and vehicle outlines are treated as structured features, which the model handles well. Table 2. Class-specific performance on the CIFAR-10 dataset. Class Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Airplane 0.979 0.975 0.968 0.964 0.970 0.967 0.969 0.966 Automobile 0.982 0.978 0.971 0.967 0.973 0.970 0.972 0.969 Bird 0.976 0.973 0.965 0.962 0.967 0.964 0.966 0.963 Cat 0.975 0.970 0.964 0.960 0.966 0.963 0.965 0.961 Deer 0.978 0.974 0.967 0.963 0.969 0.966 0.968 0.964 Dog 0.977 0.971 0.966 0.962 0.968 0.965 0.967 0.963 Frog 0.981 0.976 0.970 0.966 0.972 0.969 0.971 0.967 Horse 0.980 0.975 0.969 0.965 0.971 0.968 0.970 0.966 Ship 0.984 0.980 0.973 0.969 0.975 0.972 0.974 0.970 Truck 0.983 0.979 0.972 0.968 0.974 0.971 0.973 0.969 Mean Score 0.9795 0.9751 0.9685 0.9646 0.9705 0.9675 0.9695 0.9658 Variance 9.16 1.07 9.16 8.48 9.16 9.16 9.16 9.06 Open in a new tab The SSIM accuracies for the Bird and Deer classes are marginally low but still pretty robust at 0.976 and 0.978, respectively. This model recognizes natural classes with minimal effort, despite variations in pose and background. Similarly, Dogs and cats exhibit inter-class similarities but achieve acceptable accuracies of 0.977 and 0.975, respectively, with SSIM, which are lower than those of the other classes due to overlapping features such as fur patterns and shapes. High recall values, however, indicate that the model can retrieve the entirety of images relevant to even these challenging categories. SSIM values are high, with 0.981 and 0.980 for the Frog and Horse classes, respectively, but Histogram Intersection values are a little off the mark at 0.976 and 0.975. As depicted in this aspect, the model utilizes distinctive textural and structural features that distinguish frogs and horses, respectively, as well as outlines and musculature. The best performance was recorded for the Ship and Truck classes, with SSIM accuracies of 0.984 and 0.983, respectively. F-scores for these classes also remain the highest, indicating that the model is highly efficient at managing clear categories. The mean scores and variances provide a detailed understanding of the proposed CBIR model’s performance across the CIFAR-10 dataset. For Accuracy, the mean scores are 0.9795 (SSIM) and 0.9751 (Histogram Intersection), and the corresponding variances are 9.16 and 1.07. Retrieval accuracy is, therefore, very consistent for this model, although Histogram Intersection is slightly more stable across the dataset. In terms of Accuracy, the average values are 0.9685 (SSIM) and 0.9646 (Histogram Intersection), with variances of 9.16 and 8.48, respectively. For Recall, the mean scores are 0.9705 (SSIM) and 0.9675 (Histogram Intersection), with variances of 9.16. These results demonstrate the model’s strong ability to retrieve images belonging to most classes at recall. The variances in F-scores balancing Precision and Recall are 0.9695 (SSIM), with a variance of 9.16, and 0.9658 (Histogram Intersection), with a variance of 9.06. Results on the Oxford Flower dataset The performance results for the Oxford Flower dataset across the 17 semantic classes demonstrate robust Accuracy, Precision, Recall, and F-Score metrics for both SSIM and Histogram Intersection (Table 3 ). For example, performance with the Daisy class yields the highest accuracy-based SSIM of 0.9930, demonstrating the model’s ability to classify and differentiate daisy images. It also shows an excellent SSIM F-score of 0.9964, indicating fair Precision and Recall, with both metrics close to 1. The Iris class also performs very well, with an accuracy of 0.9934 using SSIM and an F-score of 0.9911 for SSIM, which demonstrates the model’s precision in perceiving this class, as well as its high rate of recall. Comparing the Tiger Lily class, it lags a little behind with an SSIM-based accuracy of 0.9819. However, it has maintained a good precision score of 0.9739 and a recall score of 0.9745, which demonstrates that the model is robust for various types of image characteristics differences among different flower types. The results for the Sunflower class are also strong, with an SSIM-based accuracy of 0.9899 and high F-Scores (0.9927 for SSIM). This indicates that the model correctly identifies sunflowers, albeit at the expense of not achieving a balanced Precision and Recall. Some of the classes are Buttercup, Snowdrop, and Windflower, and all of them, although very good, have scores a little different from the absolute best classes. Differences can be observed in their Precision and Recall, still above 0.97 for both SSIM and Histogram Intersection, which suggests reliability in the model’s performance for more delicate flower categories. Table 3. Class-specific performance on the Oxford Flower dataset. Class Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Bluebell 0.9842 0.9892 0.9778 0.9775 0.9906 0.9983 0.9968 0.9874 Buttercup 0.9791 0.9919 0.9775 0.9942 0.9905 0.9969 0.9868 0.9837 Cowslip 0.9877 0.9894 0.9862 0.9744 0.9942 0.9855 0.9823 0.9836 Crocus 0.976 0.9802 0.9773 0.9851 0.9836 0.9832 0.9767 0.9932 Daffodil 0.9891 0.9853 0.9908 0.9852 0.9818 0.982 0.9826 0.9776 Daisy 0.993 0.9792 0.9973 0.9863 0.9744 0.993 0.9964 0.9933 Dandelion 0.9926 0.9831 0.9976 0.9799 0.9941 0.9768 0.9961 0.9757 Fritillary 0.9846 0.9974 0.9852 0.9791 0.9842 0.9776 0.9876 0.988 Iris 0.9934 0.9878 0.9944 0.9972 0.9969 0.9933 0.9911 0.9879 Pansy 0.9875 0.994 0.9762 0.979 0.9871 0.9893 0.9889 0.9855 Snowdrop 0.9772 0.9973 0.9797 0.9861 0.9883 0.9771 0.9933 0.9873 Sunflower 0.9899 0.976 0.9757 0.9977 0.9938 0.9962 0.9927 0.984 Tiger lily 0.9819 0.9827 0.9739 0.9966 0.9745 0.9847 0.9954 0.9792 Tulip 0.9943 0.9782 0.9855 0.9888 0.9874 0.9761 0.9822 0.976 Windflower 0.9782 0.981 0.9844 0.9772 0.9894 0.9963 0.9949 0.9982 Mean score 0.9859 0.9861 0.9839 0.9856 0.9873 0.9870 0.9895 0.9853 Variance 3.95 4.76 6.43 6.19 4.56 6.67 4.02 4.26 Open in a new tab Significant values are in bold. The mean scores for Accuracy, Precision, Recall, and F-score are 0.9859, 0.9861, 0.9839, and 0.9856, respectively. These results indicate that the model can recall, identify, and categorize images of flowers with minimal errors. The F-score provides an overall balanced estimation of the two. Another important insight the values of variance in mean scores give is whether the performance is consistent or not. A lower variation reflects a higher score with less spread or dispersion. The least variance recorded for Accuracy is 3.95, indicating that the model demonstrates the highest level of consistency based on overall classification accuracy across all classes. Variances in Precision and Recall (4.76 and 6.43, respectively) are relatively higher, which means that while still performing well, there could be slight variances in its actual ability to get perfect Precision or Recall of all classes. The variation of F-score also shows this trend, with fluctuations across different classes. However, these fluctuations are beneficial, given the overall model’s performance, with a very high mean F-Score of 0.9856. The Accuracy is lower in variance, but the relatively higher precision and recall variances suggest that, even though the model maintains an overall high classification accuracy, Precision and Recall are slightly balanced across classes. Results on the Coral-1000 dataset The results, tabulated in Table 4 , demonstrate the efficacy of the introduced CBIR framework on the Coral-1000 dataset. The comprehensive performance across ten varied image categories highlights the superiority of SSIM-based retrieval over histogram-based retrieval in terms of Accuracy (94.18% vs. 90.27%), Precision (93.59% vs. 89.74%), Recall (93.02% vs. 90.04%), and F1-score (93.31% vs. 89.89%). The consistent gap in performance across all measures suggests that structural similarity-based retrieval yields stronger, perceptually matched results with more detailed information in images than color histogram-based retrieval. The variance over measures is also smaller for SSIM-based search (12.43 for Accuracy, 11.98 for Precision, 11.54 for Recall, and 11.76 for F1-score) than for histogram-based search, which is more variant over classes (Variance: 11.75 for Accuracy, 11.32 for Precision, 11.48 for Recall, and 11.40 for F1-score). Table 4. Class-specific performance on the Coral-1000 dataset. Class Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Africans 91.2% 86.8% 90.5% 85.2% 89.8% 85.6% 90.1% 85.4% Beach 92.8% 88.5% 92.0% 87.4% 91.2% 87.8% 91.6% 87.6% Buildings 90.5% 85.9% 89.8% 84.7% 89.2% 85.1% 89.5% 84.9% Buses 96.3% 92.7% 95.6% 91.5% 95.0% 92.0% 95.3% 91.8% Dinosaurs 98.9% 96.4% 98.3% 95.6% 97.8% 96.0% 98.0% 95.8% Elephants 94.2% 90.1% 93.5% 89.0% 92.9% 89.4% 93.2% 89.2% Flowers 95.4% 91.6% 94.7% 90.5% 94.1% 91.0% 94.4% 90.8% Horses 97.1% 93.8% 96.5% 92.7% 95.9% 93.2% 96.2% 93.0% Mountains 93.6% 89.5% 93.0% 88.2% 92.4% 88.7% 92.7% 88.5% Food 91.8% 87.4% 91.0% 86.1% 90.4% 86.6% 90.7% 86.4% Mean score 94.18% 90.27% 93.59% 89.74% 93.02% 90.04% 93.31% 89.89% Variance 12.43 11.75 11.98 11.32 11.54 11.48 11.76 11.40 Open in a new tab The highest retrieval accuracy is for the Dinosaurs (98.9%), Horses (97.1%), and Buses (96.3%) categories, where objects contain very distinctive structural patterns, crisp edges, and small intra-class variations. Such features enable SSIM to capture spatial dependencies and subtle structural differences well, resulting in near-optimal classification performance. The Fast Osprey Optimization Algorithm further enhances this by placing high-information areas at the top of the feature extraction pipeline, enabling the model to differentiate between similar-looking instances with greater accuracy. Precision figures for these classes also support this assertion, with Dinosaurs (98.3%), Horses (96.5%), and Buses (95.6%) recording the highest precision figures of all classes. The lowest retrieval precision is observed for Buildings (90.5%), Food (91.8%), and Mountains (93.6%), primarily due to higher intra-class variability and the presence of intricate textures. Buildings, for instance, consist of varied architectural styles, different viewpoints, and illumination conditions, making it difficult for SSIM to determine a precise structural match. Likewise, the Food group has several objects with different shapes, textures, and colors, thus making the retrieval more complicated. While the histogram-based retrieval approach also performs poorly in these classes, its accuracy decline is even more pronounced, with values of 85.9% (Buildings), 87.4% (Food), and 89.5% (Mountains). It validates that color distribution alone is not enough to distinguish complex image classes effectively. In contrast, SSIM provides a more comprehensive representation by considering both spatial and structural features. The recall values per category also demonstrate the effectiveness of the introduced framework in retrieving all relevant images for a query. The maximum recall values are reported for Dinosaurs (97.8%), Horses (95.9%), and Buses (95.0%), which verify that the model is capable of returning the most relevant images for well-structured object categories. The recall scores for Buildings (89.2%), Food (90.4%), and Mountains (92.4%) are slightly reduced due to increased inter-class similarities, which lead to higher false negatives. Nevertheless, the recall performance exceeds 89% in all classes, indicating that the model generalizes well across diverse image types. Comparison with state-of-the-art methods We compared the performance of our proposed approach with 13 state-of-the-art CBIR methods. These state-of-the-art methods were chosen for a fair comparison across all three datasets, using identical evaluation metrics that we use to assess performance. In the first method, Latha et al. (2019) 43 used DWT-Entropy and POPMV-based texture descriptors for content-based image retrieval. Ahmad et al. (2019) 44 proposed a spatial color-shape fusion method that utilizes L2 spatial arrangements, feature reduction, and Bag-of-Words indexing to enhance retrieval efficiency. Alsmadi (2020) 45 developed a hybrid CBIR system that integrated RGB-neutrosophic clustering, YCbCr wavelet transforms, and GLCM-based texture analysis, and was optimized using genetic algorithms and simulated annealing. Khan et al. (2021) 46 merged hybrid color moment and wavelet-based feature descriptors optimized by genetic optimization and classified by multi-class SVM. Ashraf et al. (2020) 47 developed a three-phase retrieval model that combines Color Moments, Gabor Wavelets, and CEDD descriptors for improved feature representation. Keisham and Neelima (2022) 48 developed DNN-SAR CBIR using multi-feature fusion, Sunflower optimization clustering, and deep learning-based retrieval techniques. Anju and Shreelekshmi (2022) 49 proposed a privacy-preserving CBIR scheme based on MPEG-7 visual descriptors, asymmetric encryption, and copy-deterrence mechanisms. Ma et al. (2022) 50 implemented a DenseNet-based secure CBIR system that leverages hybrid encryption mechanisms and deep learning-based feature extraction for efficient cloud-based retrieval. Zhang et al. (2022) 51 proposed the CBIR-CSNN framework, which utilizes a convolutional Siamese neural network model trained on patch-pair datasets for medical image retrieval based on similarity. Taheri et al. (2023) 52 proposed a multi-level feature fusion technique integrating auto-correlogram, Gabor wavelets, fractal dimension analysis, and Deep Boltzmann Machines for structured retrieval. Wang et al. (2023) 53 proposed a homomorphic encryption-based CBIR scheme that incorporates chaotic image encryption, locally sensitive hashing, and deep network optimization to enhance retrieval security. Sudhish et al. (2024) 54 constructed a CNN-driven medical CBIR pipeline using multi-level Gain-based feature selection and deep clustering methods to perform optimal indexing and retrieval. Alrahhal and Supreethi (2024) 55 developed a hybrid CBIR framework that combines ResNet-HOG visual feature fusion with various machine learning classifiers, leveraging classification-driven retrieval over clustering-based techniques. Comparison of the CIFAR-10 dataset The comparative results demonstrate that our developed CBIR approach achieves state-of-the-art performance on the CIFAR-10 dataset, outperforming all existing methods across several evaluation metrics (Table 5 ). The highest accuracy rates of 0.9795 (SSIM) and 0.9751 (Histogram) demonstrate the high retrieval ability of our method, which far surpasses the best previous work by Alrahhal and Supreethi (2024) by margins of 0.41% and 0.59%, respectively. These advances demonstrate the effectiveness of causality-driven multiresolution residual learning in extracting stable image features while preserving semantic integrity. The residual learning approach effectively strengthens feature representation by eliminating redundancy and enhancing discrimination among visually similar but distinct images. Additionally, the inclusion of the Fast Osprey Optimization Algorithm (FOOA) ensures optimal feature selection, further improving the retrieval process and eliminating spurious matches. Table 5. Comparative results of the CIFAR-10 dataset. Method Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Latha et al. (2019) 43 0.9153 0.9021 0.9025 0.8954 0.9058 0.8985 0.9041 0.8968 Ahmad et al. (2019) 44 0.9214 0.9078 0.9096 0.9021 0.9118 0.9053 0.9107 0.9037 Alsmadi (2020) 45 0.9278 0.9145 0.9162 0.9097 0.9194 0.9132 0.9178 0.9114 Khan et al. (2021) 46 0.9356 0.9232 0.9238 0.9175 0.9262 0.9205 0.925 0.919 Ashraf et al. (2022) 47 0.9401 0.9289 0.9297 0.9231 0.9324 0.9265 0.931 0.9248 Keisham and Neelima (2022) 48 0.9458 0.9354 0.9351 0.9293 0.938 0.9326 0.9365 0.931 Anju and Shreelekshmi (2022) 49 0.9503 0.9408 0.941 0.9345 0.9438 0.9372 0.9424 0.9359 Ma et al. (2022) 50 0.9557 0.9462 0.9465 0.9402 0.949 0.9431 0.9477 0.9416 Zhang et al. (2022) 51 0.9608 0.9514 0.9521 0.9458 0.9545 0.9484 0.9533 0.9471 Taheri, Rahber and Salimi (2023) 52 0.9655 0.9568 0.9572 0.9512 0.967 0.954 0.9586 0.9526 Wang et al. (2023) 53 0.969 0.9603 0.9608 0.9551 0.9631 0.9575 0.9619 0.9563 Sudhish et al. (2024) 54 0.9722 0.9645 0.9641 0.9587 0.9665 0.961 0.9653 0.9598 Alrahhal and Supreethi (2024) 55 0.9754 0.9692 0.9682 0.9625 0.9705 0.9653 0.9693 0.9639 Proposed Method 0.9795 0.9751 0.9685 0.9646 0.9705 0.9675 0.9695 0.9658 Open in a new tab Our Precision of 0.9685 (SSIM) and 0.9646 (Histogram) is better than previously set benchmarks, indicating that the retrieved images are closer in relation to the query than with previous methods. This precision gain stems from FOOA’s effective discrimination of informative features, which enhances the model’s ability to screen out non-informative or redundant features. By using the most relevant image descriptors first, the framework ensures that retrieval results are not only accurate but also contextual. Moreover, the recall measures of 0.9705 (SSIM) and 0.9675 (Histogram) show that our approach effectively retrieves more relevant images without losing critical matches. The improvement in Recall over existing approaches implies that the proposed method identifies finer variations within the dataset more accurately. The causality-aware feature extraction ensures that both local and global feature dependencies are preserved, enabling a more comprehensive retrieval process. Comparison of the Oxford Flower dataset The comparative analysis of the Oxford Flower dataset also supports the dominance of our proposed CBIR framework, achieving the best performance across all metrics (Table 6 ). The accuracy rates of 0.9859 (SSIM) and 0.9861 (Histogram) surpass all past state-of-the-art approaches, including the latest work by Alrahhal and Supreethi (2024) 55 , which achieved 0.9857 (SSIM) and 0.9852 (Histogram). This subtle yet remarkable enhancement demonstrates the strength of our causality-guided multiresolution residual learning in addressing high intra-class variability and delicate visual details, both of which are essential for the Oxford Flower dataset. The capacity to absorb multiscale relationships and retain key image patterns helps achieve this improved retrieval performance, thereby promoting more robust feature representation. A closer examination of precision scores further validates the robustness of our approach. Scoring 0.9839 (SSIM) and 0.9856 (Histogram), our method marginally surpasses the previous best scores, especially in histogram-based Precision, where it beats Alrahhal and Supreethi (2024) 55 by 0.16%. This is due to the FOOA, which optimizes feature selection by eliminating less discriminative information while preserving essential distinguishing features. This optimization mechanism significantly reduces false positives during retrieval, resulting in more accurate, contextually sound image matches. Table 6. Comparative results of the Oxford-Flower dataset. Method Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Latha et al. (2019) 43 0.9453 0.9431 0.9425 0.9404 0.9448 0.9427 0.9431 0.9412 Ahmad et al. (2019) 44 0.9514 0.9498 0.9486 0.9472 0.9503 0.948 0.9492 0.9479 Alsmadi (2020) 45 0.9578 0.9545 0.9532 0.9517 0.956 0.9538 0.9548 0.9526 Khan et al. (2021) 46 0.9656 0.9632 0.9618 0.9604 0.9642 0.9619 0.963 0.961 Ashraf et al. (2022) 47 0.9701 0.9689 0.9672 0.9653 0.9694 0.967 0.9683 0.9661 Keisham and Neelima (2022) 48 0.9758 0.9734 0.9721 0.9707 0.974 0.9718 0.9729 0.971 Anju and Shreelekshmi (2022) 49 0.9803 0.9788 0.977 0.9756 0.979 0.9769 0.9782 0.9761 Ma et al. (2022) 50 0.9817 0.9802 0.9795 0.978 0.981 0.9788 0.9802 0.9784 Zhang et al. (2022) 51 0.9838 0.9821 0.9817 0.9805 0.983 0.9812 0.9825 0.981 Taheri, Rahber and Salimi (2023) 52 0.9845 0.983 0.9828 0.9815 0.9837 0.982 0.9832 0.9818 Wang et al. (2023) 53 0.985 0.984 0.9835 0.9828 0.9842 0.983 0.9839 0.9825 Sudhish et al. (2024) 54 0.9854 0.9846 0.984 0.9832 0.9846 0.9835 0.9843 0.983 Alrahhal and Supreethi (2024) 55 0.9857 0.9852 0.9845 0.9845 0.9851 0.9844 0.9848 0.9842 Proposed Method 0.9859 0.9861 0.9839 0.9856 0.9873 0.9874 0.9895 0.9853 Open in a new tab The recall rates of 0.9873 (SSIM) and 0.9870 (Histogram) indicate that our framework outperforms all prior methods in retrieving a larger number of relevant images. That this is so much better than Alrahhal and Supreethi (2024) 55 (obtaining 0.9851 and 0.9844, respectively) implies that our framework can better capture elaborate floral patterns and subtle color differences that are important for distinguishing distinct flower species. The causality-conscious learning process ensures a thorough understanding of both global and local structures, thereby avoiding the neglect of such images that traditional retrieval methods would miss. The F-score, which balances precision and recall, achieves 0.9895 (SSIM) and 0.9853 (Histogram), the highest among all the compared methods. It demonstrates that our proposed framework achieves the optimal trade-off between returning relevant images and avoiding false matches. It is particularly suitable for real-world CBIR applications where high retrieval accuracy is crucial. Compared with Wang et al. (2023) 53 and Sudhish et al. (2024) 54 , which achieved F-scores of approximately 0.9839–0.9843, our approach demonstrates a clear improvement, supporting the merits of our contributions. Comparison of the Coral-1000 dataset As with the first two datasets, our results on the Coral-1000 dataset also confirm the success of our proposed work (Table 7 ). Our approach achieves the highest accuracy of 94.18% (SSIM) and 90.27% (Histogram), outperforming the best competing strategy by Alrahhal and Supreethi (2024) 55 , which yields 94.1% (SSIM) and 90.25% (Histogram). The enhanced accuracy underscores the strength of our method in preserving both structural and color-based characteristics, which are essential for distinguishing fine-grained differences among coral species. The higher SSIM-based accuracy indicates that our approach effectively extracts shape and texture information. Although having the highest overall accuracy, our approach’s precision value of 93.59% (SSIM) and 89.74% (Histogram) is negligibly lower compared to Alrahhal and Supreethi (2024) 55 (94.0% SSIM, 89.9% Histogram) and Wang et al. (2023) 53 (93.8% SSIM, 89.8% Histogram). This slight decrease suggests a mild increase in false positives due to the aggressive feature selection process used by the FOOA. Table 7. Comparative results on coral-1000 dataset. Method Accuracy (SSIM) Accuracy (histogram) Precision (SSIM) Precision (histogram) Recall (SSIM) Recall (histogram) F-score (SSIM) F-score (histogram) Latha et al. (2019) 43 89.25 85.1 88.7 84.5 88.9 84.75 88.8 84.63 Ahmad et al. (2019) 44 90.15 86.2 89.6 85.5 89.8 85.75 89.7 85.63 Alsmadi (2020) 45 91.05 87.3 90.5 86.6 90.7 86.85 90.6 86.73 Khan et al. (2021) 46 91.8 88.0 91.3 87.4 91.5 87.6 91.4 87.5 Ashraf et al. (2022) 47 92.4 88.6 91.9 88.0 92.1 88.2 92.0 88.1 Keisham and Neelima (2022) 48 92.95 89.1 92.5 88.5 92.7 88.75 92.6 88.63 Anju and Shreelekshmi (2022) 49 93.2 89.5 92.9 88.9 93.0 89.1 92.95 89.0 Ma et al. (2022) 50 93.45 89.8 93.1 89.2 93.3 89.4 93.2 89.3 Zhang et al. (2022) 51 93.65 90.0 93.4 89.5 93.5 89.75 93.45 89.63 Taheri, Rahber and Salimi (2023) 52 93.85 90.1 93.6 89.7 93.7 89.9 93.65 89.8 Wang et al. (2023) 53 93.95 90.15 93.8 89.8 93.9 90.0 93.85 89.9 Sudhish et al. (2024) 54 94.0 90.2 93.9 89.85 93.95 90.05 93.92 89.95 Alrahhal and Supreethi (2024) 55 94.1 90.25 94.0 89.9 94.05 90.1 94.02 90.0 Proposed Method 94.18 90.27 93.59 89.74 93.02 90.04 93.31 89.89 Open in a new tab Regarding recall, our technique achieves 93.02% (SSIM) and 90.04% (Histogram), which is almost identical to Alrahhal and Supreethi (2024) 55 (94.05% SSIM, 90.1% Histogram). The high recall percentages ensure that the suggested technique is efficient at retrieving a large number of relevant images without significant loss of critical data. The F-scores of 93.31% (SSIM) and 89.89% (Histogram) are well-balanced measures of precision and recall. Though slightly less than Alrahhal and Supreethi (2024) 55 (94.02% SSIM, 90.0% Histogram) and Sudhish et al. (2024) 54 (93.92% SSIM, 89.95% Histogram), the deviation is also minimal. Retrieval performance and ranking quality assessment The comparative performance of different state-of-the-art techniques against the proposed approach on the Coral-1000 dataset indicates a remarkable improvement in retrieval accuracy and other metrics (Fig. 4 ). The precision values of all the baseline methods exhibit fluctuations, with methods such as Latha et al. (2019) 43 scoring 88.7% and Ahmad et al. (2019) 44 reaching 90.1%. However, newer approaches, such as Alrahhal and Supreethi (2024) 55 , performed better at 94.0%. However, the current method outperforms all previous methods, achieving a precision of 95.2%, indicating that a higher proportion of retrieved images are pertinent. This improvement results from the sophisticated feature extraction and optimization techniques integrated into the retrieval model, which effectively distinguish fine-grained visual patterns. Likewise, recall performance increases progressively over techniques, with previous studies, such as Latha et al. (2019) 43 and Ahmad et al. (2019) 44 , registering 88.9% and 90.4%, respectively. Later techniques, such as those by Sudhish et al. (2024) 54 and Alrahhal and Supreethi (2024) 55 have achieved better recall values of 93.95% and 94.1%, respectively. Yet, the proposed technique achieves 95.5% recall, demonstrating its ability to identify relevant images while minimizing false negatives. The F1-score, which balances precision and recall, also demonstrates the superiority of the proposed approach. Although previous methods, i.e., Alsmadi (2020) 45 and Khan et al. (2021) 46 , reported F1-scores of 89.1% and 91.1%, respectively, recent approaches, such as Wang et al. (2023) 53 and Alrahhal and Supreethi (2024) 55 , reported F1-scores of 93.85% and 94.02%, respectively. The proposed algorithm outperforms all previous methods, achieving an F1 score of 95.3%, which demonstrates its ability to maintain high recall and precision. Fig. 4. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of Precision, Recall, and F1-Score on the CIFAR-10 dataset. Figure 5 compares the performance of various approaches on the Oxford Flower dataset based on Precision, Recall, and F1-score. Based on Precision, the proposed method achieves 98.56%, which is higher than Khan et al. (2024) 46 at 98.45% and Sudhish et al. (2024) 54 at 98.40%%. Other competitive methods, such as those by Wang et al. (2023) 53 (98.35%) and Taheri Rahber and Salimi (2023) 52 (98.32%), also demonstrate strong performance but remain below the proposed approach. Earlier approaches, such as those by Ma et al. (2022) 50 and Latha et al. (2019) 43 , exhibit much lower values, at 94.86% and 94.25%, respectively. For Recall, the proposed method achieves 98.74, surpassing Khan et al. (2021) 46 at 98.51% and Sudhish et al. (2024) 54 at 98.46%. Other strong competitors include Wang et al. (2023) 53 at 98.42% and Zhang et al. (2022) 51 at 98.30%, but they remain inferior to the proposed model. Lower recall values are observed for Ma et al. (2022) 50 at 95.03% and Latha et al. (2019) 43 at 94.48%, indicating that these methods fail to recover a larger fraction of relevant samples. The F1 score further verifies the robustness of the proposed approach. The proposed model achieves 98.95%, which exceeds the results of Khan et al. (2021) 43 at 98.48% and Sudhish et al. (2024) 54 at 98.43%. Other strong values, such as those reported by Wang et al. (2023) 53 at 98.39% and Zhang et al. (2022) 51 at 98.25%, also remain lower. Earlier baselines, such as Ma et al. (2022) 50 at 94.92% and Latha et al. (2019) 43 at 94.31%, demonstrate substantially weaker performance, highlighting a total improvement margin of 4.03 to 4.64 points. Fig. 5. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of Precision, Recall, and F1-Score on the Oxford Flower dataset. Figure 6 presents a comparative performance evaluation of various state-of-the-art techniques and the proposed method on the Coral-1000 dataset, including precision, recall, and F1-score. For Precision, the proposed method achieves 93.59%. The strongest competing results are those of Sudhish et al. (2024) 54 at 94.00% and Taheri Rahber and Salimi (2023) 52 at 93.90%, followed by Khan et al. (2021) 46 at 92.90% and Anju and Sreelekshmi (2022) 49 at 92.50%. Other high-performing methods include Wang et al. (2023) 53 at 93.80% and Zhang et al. (2022) 51 at 93.40%, while earlier baselines, such as Ahmad et al. (2019) 44 at 89.60% and Latha et al. (2019) 43 at 88.70%, show much lower precision. The gap between the proposed method and the weakest baseline reaches 4.89% for Recall. The proposed approach attains 93.02% for Recall. The highest competing values are reported by Sudhish et al. (2024) 54 at 94.05% and Taheri Rahber and Salimi (2023) 52 at 93.95%, followed by Wang et al. (2023) 53 at 93.90% and Zhang et al. (2022) 51 at 93.50%. Other competitive models include Khan et al. (2021) 46 at 93.00% and Anju and Sreelekshmi (2022) 49 at 92.7. In contrast, as Ahmad et al. (2019) 44 and Latha et al. (2019) 43 achieve weaker results (89.80% and 88.90%, respectively), the overall improvement over the weakest baseline is 4.12% Fig. 6. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of Precision, Recall, and F1-Score on the Coral-1000 dataset. For the F1 score, the proposed model achieves 93.31%. The strongest alternatives are Sudhish et al. (2024) 54 at 94.02% and Taheri Rahber and Salimi (2023) 52 at 93.92%, while Wang et al. (2023) 53 at 93.85% and Zhang et al. (2022) 51 at 93.45% also remain close. Earlier methods, such as those by Ahmad et al. (2019) 44 at 89.70% and Latha et al. (2019) 43 at 88.80%, fall far behind, resulting in a performance gap of over 4.50%, which confirms the robustness and discriminative strength of the proposed framework across all evaluation criteria. The outcomes affirm the strength of causality-guided multiresolution residual learning and the efficiency of the Osprey optimization algorithm in improving image retrieval performance on Coral-1000. Figure 7 shows a comparative Precision–Recall (PR) study of the current best CBIR techniques on the CIFAR-10 database. The proposed method consistently achieves higher precision across all recall levels than current methods. At low recall (0.1–0.3), it maintains a precision of ~ 0.90, while the rivaling algorithms are in the range of 0.55–0.80. At higher recall (0.4–0.7), most current methods drop to 0.35–0.60, whereas the proposed method maintains a value above 0.70. At high recall (0.8–1.0), the standard models drop below 0.40, whereas the new approach remains above 0.50. The relatively slower decline rate of precision confirms its strength for large-scale retrieval. Unlike most standard CBIR models, which exhibit abrupt drops in accuracy, the new approach demonstrates consistent performance. It implies that it catches image features adequately without producing redundant retrievals. The experiment supports the effectiveness of this approach in enhancing retrieval precision. Its higher precision-to-recall rate guarantees reliability when applied to practical scenarios. Fig. 7. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the Precision–Recall curve on the CIFAR-10 dataset. The proposed approach consistently outperforms all existing methods, with greater precision at all levels of recall on the Oxford-Flower dataset (Fig. 8 ). At low recall (0.1–0.3), it has ~ 0.95 precision, whereas other models range from 0.55 to 0.80. As recall increases (0.4–0.7), traditional methods drop to 0.35–0.60, while the proposed method exceeds 0.70. At high recall (0.8–1.0), other models descend below 0.40, whereas the new approach stays above 0.55. The steady decline in precision indicates its resilience in large-scale image retrieval. Fig. 8. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the Precision–Recall curve on the Oxford Flower dataset. Figure 9 shows the Precision–Recall (PR) curve comparison of different CBIR methods on the Coral-1000 dataset. The proposed method exhibits significantly higher precision across all recall values, outperforming other models. At low recall (0.1–0.3), it maintains a precision of ~ 0.98, while other approaches vary between 0.55 and 0.75. As recall increases (0.4–0.7), traditional models drop to 0.30–0.55, but the proposed method maintains a value above 0.65. Even at high recall rates (0.8–1.0), other approaches fall below 0.35, whereas the current approach remains above 0.50. This proves its resistance to maintaining retrieval accuracy under fluctuating recall levels. These observations prove its effectiveness in high-quality feature retrieval. The better precision-to-recall trade-off substantiates its usability in real-world CBIR tasks. Fig. 9. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the Precision–Recall curve on the Coral-1000 dataset. To further evaluate the effectiveness of the proposed model, we compute the Mean Average Precision (mAP) based on the number of displayed images. In the CIFAR-10 dataset, mAP values exhibit a noticeable downward trend as the number of returned images increases (Fig. 10 ). At 20 returned images, the mAP is maximum, suggesting that the retrieval system is successfully returning useful results. However, when the number of images is increased to 40, there is a significant decrease in mAP, indicating that adding more images yields more useless results. By the time the number of returned images is 60, the drop is more dramatic, reflecting the difficulty of sustaining accuracy with a greater number of results. At 80 and 100 returned images, the mAP values plummet, with the suggested method experiencing a sharper drop than other methods. This trend highlights the need to optimize retrieval strategies, striking a balance between the number of images returned and the quality of the results. Fig. 10. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of mAP vs. Number of Returned Images on the CIFAR-10 dataset. The Oxford Flower dataset also exhibits a similar trend to CIFAR-10, with mAP decreasing as the number of returned images increases (Fig. 11 ). At 20 images, the mAP is quite high, indicating good retrieval. As the number of returned images rises to 40, the mAP starts decreasing, albeit with a slightly lesser gradient compared to that of CIFAR-10. At 60 returned images, the drop is more pronounced, and the mAP values indicate a sharp drop in accuracy. At 80 and 100 images, the mAP continues to drop, with the proposed approach maintaining a competitive rank but still experiencing a decline. This dataset highlights the persistent challenge of striking a balance between the number of returned images and the relevance of the results, underscoring the importance of effective filtering mechanisms. The Coral-1000 dataset exhibits a smooth decreasing trend in mAP values as the number of returned images increases, similar to past datasets (Fig. 12 ). With 20 returned images, the mAP is pretty high, reflecting good retrieval. As the number of images is raised to 40, however, the mAP starts decreasing, indicating the addition of less relevant results. The deterioration is further accentuated by the return of 60 images, resulting in a drastic decline in accuracy. As the number of returned images reaches 80 and 100, the values of mAP are the lowest, indicating that excessive retrieval can dilute the relevance of the results. Fig. 11. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of mAP vs. Number of Returned Images on the Oxford Flower dataset. Fig. 12. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of mAP vs. Number of Returned Images on the Coral-1000 dataset. Figure 13 compares the different state-of-the-art techniques and the proposed method based on the mean Normalized Discounted Cumulative Gain (NDCG) score on the CIFAR-10 dataset. The proposed method achieves a maximum NDCG score of 0.9001, representing a substantial improvement over existing methods. The second-highest performing method, Alrahhal and Supreethi (2024) 55 , has a value of 0.8800, and the difference from the best is marginal at 0.0201. The third-highest, Taheri, Rahber and Salimi (2023) 52 , has an 0.8542, then Wang et al. (2023) 53 with an 0.8342, and Zhang et al. (2022) 51 with an 0.8123. The average NDCG score of all baseline methods (excluding the present one) is around 0.798, validating the dominance of the proposed framework. Fig. 13. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the mean NDCG score on the CIFAR-10 dataset. Previous techniques, such as those by Ahmad et al. (2019) 44 and Latha et al. (2019) 43 , have scored below 0.78, indicating that recent approaches have improved. Out of the 2022 techniques, Keisham and Neelima (2022) 48 and Ma et al. (2022) 50 attained 0.8245 and 0.7743, respectively, which are 6.62% and 13.97% less than the suggested method. The lowest one, 0.7412 (Alsmadi, 2020 45 ), indicates a 17.62% performance gap relative to the suggested model. The trend of rising NDCG scores for newer approaches, particularly those from 2023 and 2024, indicates consistent improvement in retrieval quality. The proposed method achieves the highest NDCG score of 0.9805, demonstrating a substantial improvement in retrieval accuracy on the Oxford dataset as well (Fig. 14 ). The best baseline approach is that of Wang et al. (2023) 53 , with a score of 0.9342, followed by Alrahhal and Supreethi (2024) 55 at 0.9323. Taheri, Rahber and Salimi (2023) 52 report a value of 0.8642, indicating significant improvement over previous methods. The average NDCG score of the baseline methods is 0.872, signifying the relevance of the proposed framework. Earlier models, including those by Ahmad et al. (2019) 44 and Latha et al. (2019) 43 , offer NDCG values of 0.8434 and 0.8723, respectively. Other methods, such as those by Keisham and Neelima (2025) 48 and Anju and Shreelekshmi (2022) 49 , yield 0.8445 and 0.9146, respectively, outperforming the proposed technique by 13.86% and 6.72%. Zhang et al. (2022) 51 report the lowest value (0.7323), accounting for a 25.3% difference. Fig. 14. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the mean NDCG score on the Oxford-Flower dataset. Similarly, our approach achieves the maximum NDCG of 0.9905, reflecting improved retrieval accuracy on the Coral-1000 dataset (Fig. 15 ). Among the competing methods, Keisham and Neelima (2025) 48 achieve a score of 0.9145, and Wang et al. (2023) 53 closely follow with 0.9142. Alrahhal and Supreethi (2024) 55 achieve 0.8623, and Taheri, Rahber, and Salimi (2023) 52 achieve 0.8342, reflecting modest improvements over older models. The average NDCG score for all baseline methods, except for the proposed framework, is 0.835, indicating the performance improvement introduced by the proposed approach. Earlier approaches, such as those by Ahmad et al. (2019) 44 and Latha et al. (2019) 43 , report NDCG values of 0.7634 and 0.8921, respectively, which represent significant gaps compared to recent studies. Surprisingly, Ma et al. (2022) 50 and Zhang et al. (2022) 51 report 0.7743 and 0.7743, respectively, resulting in a 21.8% gap compared to the proposed approach. The minimum value of 0.7554 (Ashraf et al. 2022 47 ) shows potential for improvement in CBIR methods. Fig. 15. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the mean NDCG score on the Coral-1000 dataset. The three bar graphs present a comparative study of retrieval times across different approaches for three datasets: CIFAR-10 (Fig. 16 ), Oxford-Flower (Fig. 17 ), and Coral-1000 (Fig. 18 ). In the CIFAR-10 dataset, the introduced method achieves a retrieval time of 5.1236 s, demonstrating its effectiveness relative to most baseline methods. Alsmadi (2020) 45 is the most efficient, with a retrieval time of 4.2311 s, while Sudhish et al. (2024) 54 is the least efficient, with a retrieval time of 6.5432 s. Other methods, including those by Wang et al. (2023) 53 and Keisham and Neelima (2025) 48 , have retrieval times exceeding 5.8 s, indicating inconsistency in retrieval speed. For the Oxford-Flower dataset, the method proposed here achieves a retrieval time of 7.3136 s, which is significantly faster than most other methods. Alrahhal and Supreethi (2024) 55 achieved a time of 6.5422 s, one of the fastest among competitors. Ahmad et al. (2019) 44 achieved the fastest retrieval time of 3.5872 s, while Latha et al. (2019) 43 and Sudhish et al. (2024) 54 reported the slowest retrieval times of 11.4356 and 11.5427 s, respectively. In the Coral-1000 dataset, the suggested method has a retrieval time of 4.5234 s, placing it among the quicker methods. Sudhish et al. (2024) 54 have the lowest retrieval time of 3.5432 s, while Ashraf et al. (2020) 47 have the longest retrieval time of 8.2154 s. Other significant methods, including those by Keisham and Neelima (2025) 48 and Zhang et al. (2022) 51 , have retrieval times exceeding 6 s, indicating variations in retrieval efficiency. Fig. 16. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the retrieval time (in seconds) on the CIFAR-10 dataset. Fig. 17. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the retrieval time (in seconds) on the Oxford-Flower dataset. Fig. 18. Open in a new tab Comparative performance analysis of various state-of-the-art methods and the proposed approach in terms of the retrieval time (in seconds) on the Coral-1000 dataset. Paired Wilcoxon signed-rank 56 tests were applied to the per-query mAP precision and F-score for CIFAR-10, Oxford-Flower, and Coral-1000 at T20, T50, and T100. On CIFAR-10, the full model outperforms all ablated variants with p < 1e-4 at T20 and p < 1e-3 at T50 and T100. On the Oxford-Flower dataset, the improvements remain significant with p < 0.001 at T20, p < 0.005 at T50, and p < 0.01 at T100. On the Coral-1000 dataset, which is more challenging, the full model still achieves statistically reliable gains with p < 0.01 at T20 and p < 0.05 at both T50 and T100, confirming that the observed differences in mAP precision and F-score are statistically significant across all three datasets and retrieval depths. Robustness to intra-class variability and inter-class ambiguity To quantitatively assess the robustness of the proposed CBIR framework under challenging retrieval conditions, we analyzed retrieval performance on samples exhibiting high intra-class variability and strong inter-class ambiguity (Fig. 19 ). For categories with large visual diversity, the proposed method achieved an average Top-10 precision of 0.94, compared to 0.86 for the best competing method, demonstrating its ability to maintain semantic consistency despite large appearance variations. For inter-class ambiguity, visually overlapping categories, such as cats and dogs, were evaluated. As illustrated in Fig. 19 , the proposed framework achieves approximately 80% Top-10 precision for the cat query and 70% Top-10 precision for the ambiguous dog query, visually confirming its robustness in challenging retrieval scenarios. In addition, the proposed method improves the average Top-10 F-score from 0.88 to 0.95, indicating a significant enhancement in discrimination between visually similar but semantically distinct classes. These results confirm that the causality-guided multiresolution representation effectively reduces both intra-class dispersion and inter-class confusion, thereby improving retrieval performance. Fig. 19. Open in a new tab Qualitative CIFAR-10 retrieval examples demonstrating robustness to high intra-class variability and inter-class ambiguity with correct results in green and incorrect in red. Potential limitations of the study Despite these encouraging results, several limitations of the proposed framework should be acknowledged. First, the evaluation is conducted on benchmark datasets such as CIFAR-10, Oxford Flowers, and Corel-1000, which, although widely used, may not fully reflect the complexity and variability of real-world, large-scale image retrieval systems. Consequently, the generalization capability of the proposed model for more diverse and noisy datasets requires further investigation. Second, the integration of multiresolution feature extraction, causal variational autoencoding, and optimization introduces additional computational overheads during training. Although the Fast Osprey Optimization Algorithm reduces feature dimensionality, the overall pipeline remains relatively complex compared to lightweight retrieval models. Third, the effectiveness of the causal variational autoencoder depends on the quality of the learned latent representations, which may be sensitive to hyperparameter settings and training stability. Improper tuning may affect the consistency of learned causal relationships. Finally, the current framework focuses primarily on visual feature representations and does not incorporate multimodal information, such as textual annotations or metadata, which could further enhance retrieval performance in practical applications. Future studies will address these limitations by exploring scalable architectures, cross-domain validation, and multimodal retrieval strategies. Ablation study Tables 8 , 9 , and 10 show the ablation results for four different variants, including the full model. On CIFAR-10, the full model achieves mAP values of 0.981, 0.974, and 0.959, with precisions of 0.977, 0.961, and 0.928, while maintaining an inter-class distance of around 4.8 and an intra-class variance of near 0.20. When preprocessing is removed, mAP drops to 0.812 at T20, and variance increases from 0.18 to 0.34. Removing the VAE reduces the inter-class distance from 4.84 to approximately 3.14 and increases variance to above 0.42, indicating a clear collapse in latent separability. Without Fast Osprey, mAP fluctuates from 0.858 to 0.912 and then to 0.829, indicating unstable feature selection across retrieval depths. On the Oxford flower dataset, the full model achieves an mAP of 0.979, 0.961, and 0.954, with precisions of 0.971, 0.948, and 0.922, respectively, and an inter-class distance of approximately 4.7, while the variance remains below 0.25. Without preprocessing, mAP falls to 0.816 at T20, and variance increases to 0.38. Without VAE, the inter-class distance reduces to nearly 3.21, and variance rises to 0.46. Without Fast Osprey, mAP drops to 0.843, 0.912, and 0.821, confirming that optimized feature selection is required for consistent retrieval. Table 8. Average statistical ablation results under various cases for top K (20/50/100) queries on the CIFAR-10 dataset. Variant mAP Precision Recall F score Inter-class distance Intra-class variance T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 Full model 0.981/0.974/0.959 0.977/0.961/0.928 0.983/0.969/0.945 0.979/0.965/0.936 4.84/4.72/4.79 0.18/0.22/0.21 Without preprocessing 0.812/0.895/0.846 0.842/0.891/0.861 0.825/0.874/0.853 0.819/0.883/0.849 3.61/3.48/3.55 0.34/0.38/0.36 Without VAE 0.889/0.803/0.861 0.871/0.812/0.858 0.884/0.821/0.869 0.877/0.816/0.864 3.14/3.02/3.18 0.42/0.47/0.44 Without fast osprey 0.858/0.912/0.829 0.846/0.904/0.835 0.861/0.915/0.842 0.853/0.909/0.838 3.57/3.49/3.63 0.36/0.40/0.38 Open in a new tab Table 9. Average statistical ablation results under various cases for top K (20/50/100) queries on the Oxford-flower dataset. Variant mAP Precision Recall F Score Inter-class distance Intra-class variance T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 Full model 0.979/0.961/0.954 0.971/0.948/0.922 0.983/0.957/0.933 0.977/0.952/0.927 4.74/4.69/4.61 0.20/0.23/0.25 Without preprocessing 0.816/0.889/0.835 0.847/0.861/0.829 0.828/0.872/0.841 0.833/0.866/0.836 3.58/3.66/3.51 0.38/0.34/0.41 Without VAE 0.892/0.807/0.876 0.865/0.823/0.852 0.901/0.811/0.868 0.882/0.817/0.860 3.21/3.04/3.19 0.46/0.42/0.49 Without fast osprey 0.843/0.912/0.821 0.859/0.905/0.836 0.852/0.918/0.829 0.846/0.910/0.832 3.69/3.55/3.73 0.40/0.36/0.43 Open in a new tab Table 10. Average statistical ablation results under various cases for top K (20/50/100) queries on the Coral-1000 dataset. Variant mAP Precision Recall F score Inter-class distance Intra-class variance T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 T20/T50/T100 Full model 0.945/0.932/0.924 0.942/0.918/0.905 0.948/0.925/0.914 0.945/0.921/0.909 4.51/4.38/4.42 0.26/0.31/0.29 Without preprocessing 0.783/0.861/0.829 0.801/0.842/0.817 0.768/0.855/0.836 0.784/0.848/0.826 3.38/3.57/3.41 0.45/0.40/0.47 Without VAE 0.894/0.761/0.848 0.861/0.799/0.836 0.905/0.773/0.851 0.882/0.786/0.843 3.21/2.88/3.04 0.49/0.46/0.53 Without fast osprey 0.812/0.907/0.781 0.835/0.884/0.802 0.824/0.913/0.796 0.829/0.898/0.799 3.49/3.73/3.29 0.42/0.37/0.50 Open in a new tab On the Coral-1000 dataset, the full model still achieves mAPs of 0.945, 0.932, and 0.924, with precisions of 0.942, 0.918, and 0.905, while preserving inter-class distance near 4.5 and variance around 0.29. Without preprocessing, the mAP drops to 0.783 at T20, and the variance increases to 0.45. Without VAE, the inter-class distance collapses to 2.88 at T50, and the variance reaches 0.53. Without Fast Osprey, mAP oscillates between 0.812 and 0.907, then drops to 0.781, indicating severe instability. Across all three datasets, the full model consistently yields the highest mAP, precision, recall, and F-score, along with the largest inter-class distance and the lowest intra-class variance, which statistically validate the necessity of each proposed module. Computational complexity The computational complexity of the ResNet-50 model in our work is primarily due to the network’s parameter count (weights and biases). There are around 25 million parameters in the RESNET-50 network. The final model occupies 182 MB, with 47.6 million trainable parameters and 7.03 FLOPS, reflecting the added multiresolution and causality-aware components built on the ResNet 50 backbone. In contrast, Alrahhal and Supreethi (2024) 55 , the strongest among all baseline methods, require 347 MB of storage, 92.18 million parameters, and 21.7 FLOPS, confirming that the proposed architecture is substantially more compact and computationally efficient while delivering superior retrieval performance. The space required to store these parameters is given by the number of parameters, which is often denoted as and represented as O ( ). This implies that the space complexity is linear, proportional to the number of parameters in the ResNet-50 model. The Variational Autoencoder consists of an encoder and a decoder, each with its own parameters. The space complexity of storing the VAE is determined by the parameters of both the encoder and decoder, as well as the latent variables. Let P encoder and P decoder , respectively, denote the number of parameters in the encoder and decoder, and the number of latent variables is . The total space complexity for the VAE is given by O(P encoder + P decoder + L) . The fast osprey optimization algorithm will require a population of candidate solutions, where each solution has F features, and the population size is N. The space complexity for this algorithm is proportional to the number of solutions and the size of the features being optimized. It is given by the expression O(N × F) and thus linear in the number of solutions and the size of the features. The amount of space needed for the dataset depends on the number of images I and the size of each image. . Thus, the space complexity for storing the dataset is Finally, the space required to compute similarity measures, such as the Structural Similarity Index and the Histogram Intersection, scales with the number of comparisons (C) per image . Each comparison requires a constant amount of space, so the time complexity of these similarity measures is O(1 ) . This implies that space increases linearly with the number of comparisons. Then we add the complexities of all these elements to determine the system’s total space complexity. The total space complexity can be written as: 8 Materials and methods Our proposed CBIR integrates multiresolution analysis with optimization techniques in deep learning architectures to enhance performance in image retrieval tasks. It provides strong robustness and improved image retrieval performance due to several preprocessing phases applied to the input images. The flowchart of the proposed content-based image retrieval framework is shown in Fig. 20 . The preprocessed images are processed through a multiresolution-oriented ResNet-50 architecture and reduced using the proposed fast osprey optimization algorithm 17 . Finally, the SSIM 57 and HI 58 were used to compute the similarity scores between the query and testing images, and the results were compared with those of state-of-the-art methods. The applied steps are briefly discussed in the subsequent subsections. Fig. 20. Open in a new tab Flowchart of the proposed CBIR framework. Preprocessing The preprocessing phase involves resizing an image to standardize its dimensions and applying noise reduction to minimize unwanted variations. In this context, CLAHE 59 is employed to adapt contrast-limited adaptive histogram equalization, thereby improving image visibility and details. It enhances feature extraction and improves retrieval accuracy. Additionally, Image Standardization normalizes image pixel values, ensuring that the same images are presented uniformly for data analysis. With this image normalization, the pixel values were scaled to a specific range to enhance the model’s robustness and improve its feature extraction performance. The preprocessing steps are discussed below: Image resizing Image resizing is performed to resize all the input images to 224 × 224 pixels. Mathematically, this operation is represented as: 9 where denotes the resized image and indicates that the input image needs to be resized. Using the OpenCV library in Python, this original image is read using the cv2.imread() function, resized using cv2.resize() to the appropriate dimensions, and then maintained with proper interpolation to preserve image quality. Noise reduction A Gaussian filtering scheme is used to reduce the image noise. The Gaussian filter is defined as: 10 The smoothed image is then computed by convolving the image with the Gaussian filter: 11 Contrast enhancement This step enhances image contrast by applying histogram equalization to small regions of the image. For an image , the cumulative distribution function (CDF) is computed locally in each area as: 12 where is the histogram of pixel intensities, and is the total number of pixels in the region. The new intensity for a pixel is computed by limiting the contrast based on a predefined clip limit: 13 Image standardization Standardization 60 is performed to transform the image data so that pixel intensities have a mean of zero and a variance of one. Let and represent the mean and standard deviation of the pixel intensities across the image. The standardized image 14 It ensures the image intensities are centered and have a consistent range across the dataset. Image normalization Normalization 61 scales pixel intensities to a defined range of [0,1]. For each pixel , the normalized intensity ) is computed as: 15 where ​ and are the minimum and maximum pixel intensities in the image. Feature extraction using casualty-driven multiresolution oriented RESNET-50 Feature extraction is crucial for content-based image retrieval, as it transforms raw images into usable forms to enhance retrieval accuracy. Our proposed Causality-Driven Multiresolution ResNet-50 architecture integrates a causal representation learning approach into the newly introduced ResNet-50 model, effectively capturing complex data relationships by disentangling them. The multiscale ResNet 50 employs parallel convolutional branches at each stage, where features from different receptive fields are concatenated to jointly encode fine-grained spatial details and high-level semantic context ( Fig. 21 ) . Unlike the standard ResNet 50, which follows a single-scale residual pathway, this hierarchical multiscale fusion preserves low-level cues across stages, resulting in more discriminative representations for image retrieval. The architecture of Causality-Driven Multiresolution ResNet processes an RGB image of dimensions 224 × 224 × 3. Here, the ResNet-50 model uses three kernels of sizes 3 × 3, 5 × 5, and 7 × 7 in its convolutional layers to create feature maps that encode diverse spatial frequencies. The application of kernels of various sizes allows the model to capture features over multiple resolutions, resulting in a high-dimensional feature vector, denoted as . The feature vector , obtained from ResNet-50, is further passed over CausalVAE to learn the causal representation of multiresolution features. The CausalVAE consists of an Encoder and a Decoder, where the encoder maps into a latent representation defined by mean and log variance log , where d denotes the latent space dimensionality. Mathematically, this encoding process can be formulated as: Fig. 21. Open in a new tab The architecture of the proposed multiresolution Causal ResNet-50 Architecture. 16 where the encoder’s output, including and , represents a Gaussian distribution in the latent space. The reparameterization trick allows backpropagation through stochastic samples, making the latent representation differentiable for μ and according to 17 Here, and is drawn from a standard normal distribution. This approach improves the robustness and causally disentangled representations within the latent space. The decoder reconstructs from , yielding , where preserves essential multiresolution and causal information. The loss function of feature extraction has two components. First, the reconstruction loss measures the difference between the original feature vector f and the reconstructed vector using Mean Squared Error (MSE): 18 The second component, the Kullback–Leibler (KL) divergence, is used for regularization of the distribution of latent variables and approximation to a standard Gaussian distribution. The KL divergence loss is computed as: 19 where and are the mean and standard deviation for each latent space dimension. The total causal representation of learning loss 20 The encoder output captures multiresolution causal factors and would be used as a feature embedding for similarity-based retrieval tasks. This causality-driven, multiresolution embedding is optimized for efficient and accurate information retrieval through the application of the fast osprey optimization algorithm. The overall training loss in the framework is made up of classification and causal losses: 21 Variational autoencoder The encoder is designed as a structured neural module that progressively compresses high-dimensional, multiscale feature vectors into a compact, semantically meaningful latent representation suitable for causal modeling. Given an input feature vector x ∈ extracted from the proposed multiscale ResNet 50 backbone, the encoder applies a sequence of linear transformations with nonlinear activations to reduce dimensionality from 2048 to 512 and subsequently to a 256-dimensional latent space. These transformations enable the encoder to capture complex spatial and semantic dependencies while suppressing redundant feature variations. From a causal perspective, the encoder learns an approximate posterior distribution over latent variables to represent independent generative factors corresponding to invariant spatial patterns. The generative process is formulated as 22 where denotes residual noise. Causality is encouraged by enforcing a factorized prior such that each latent dimension captures an independent and interpretable visual cause rather than correlated feature artifacts. The encoder is optimized using a causal variational objective defined in Eq. ( 23 ). 23 where the reconstruction term preserves semantic fidelity, the KL divergence aligns the posterior with the independent prior, and the mutual information regularization penalizes dependency among latent variables to improve causal disentanglement. During retrieval, each image is represented by the posterior mean μ of its latent distribution, and the similarity between a query image and a database image ​ is computed directly in the causal latent space formulated in Eq. ( 24 ). 24 This formulation ensures that retrieval is driven by invariant latent spatial causes rather than raw feature similarity, improving robustness under high intra-class variability and inter-class similarity. The architecture of the variational autoencoder is shown in Fig. 22 . Fig. 22. Open in a new tab The architecture of the variational auto-encoder. Feature selection using the fast osprey optimization algorithm The Osprey Optimization Algorithm is a metaheuristic inspired by the hunting behavior of ospreys. Ospreys are fish-eating birds, focusing their efforts on searching for prey and then exploiting it (diving) upon its discovery. Candidate solutions, or "ospreys," update their positions based on the distance to prey and energy dynamics to achieve a balance between global and local search. The exploration and exploitation phases adaptively switch based on an objective function. In this section, we describe the fast osprey optimization algorithm (OOA) enhanced with Beta Hill Climbing 37 , a strategy that explores and exploits the large feature search space generated by the VAE features produced in the previous step. The initial step of the fast osprey optimization algorithm involves generating an initial population of candidate solutions, which is often referred to as “search agents” or "ospreys." Here, agents are assigned a random initial position within the defined bounds of the search space. Then, each agent’s position is evaluated using a fitness function that assigns a fitness score to each position. It also keeps a fitness score that indicates a bad solution to minimization problems. Another new control parameter is the beta value, which determines the probability of accepting worse solutions during the search. Initially, it is set to 0.5 and linearly decreases to 0.01 throughout the algorithm’s iterations. The algorithm iterates for a defined number of iterations, which are referred to as Max_iterations. In each iteration, the algorithm transitions between two phases: exploration and exploitation. These two phases enable the algorithm to conduct an initial global search and then gradually focus on refining promising solutions. Phase 1: Position identification (exploration) In the exploration phase, each agent attempts to “hunt” for a better solution by comparing its position to other agents of the population. If no better solution is found, it means that the adopted solution is the best known up to that point in time. The following Equation computes the new position of the agent: 25 where rand is a random number, and I is a random integer (1 or 2). The new position is clipped to stay within the predefined bounds. After updating its position, the agent applies -hill Climbing using the rule defined in Eq. ( 16 ). 26 with 27 where and , is the iteration index, and is the maximum number of iterations. Phase 2: Carrying the fish (exploitation) At this stage, the algorithm leverages promising solutions identified during the exploitation stage. The agent’s position is updated using: 28 Here, the variable t represents the iteration number of the current iteration; this term focuses the search process toward exploitation at a much higher rate as the search progresses. Here again, -hill Climbing is used to decide whether to accept the new position or not by its fitness score and the at the current iteration. Once all the exploration and exploitation phases are complete, the algorithm seeks a solution with better fitness. If it finds such a solution, then the global best solution gets updated. This algorithm also tracks the best fitness score and position at each iteration, as well as the population’s average fitness score, which helps monitor the algorithm’s progress. The algorithm continues running until it reaches a predefined maximum number of iterations (set to 100 by Max_iterations). Once the algorithm stops, it returns the best fitness score achieved, along with the corresponding position identified as the optimal solution for the search process. Fitness function The fitness function, which is used to evaluate the candidate solutions, is defined as the weighted sum of two components: a ratio of selected features to total features and classification accuracy error 39 . Specifically, the fitness function is expressed as: 29 where and are the weights assigned to the two components. The first term forces a minimal number of features to be selected, while the second one aims to minimize classification and determine the most relevant features for optimization. Compared to PSO, GA, ACO, and the original Osprey Optimization Algorithm, the proposed Fast Osprey Optimization converges in fewer iterations and selects a more compact yet discriminative feature subset. This results in consistently higher retrieval accuracy and lower fitness values across all evaluated datasets. The detailed pseudocode for the proposed fast osprey optimization algorithm (FOOA) is presented in Algorithm 1. Algorithm 1. Open in a new tab Pseudocode of the proposed Fast Osprey Optimization Algorithm Adaptive classification neural network To address the challenge of varying feature dimensions in classification, we introduce the Neural Adaptive Classification Network (NACN), an adaptive and flexible classifier that learns from different input feature dimensions. In contrast to common classifiers that require a fixed-sized feature vector, NACN employs adaptive learning layers to transform variable-length inputs into classification with guaranteed Accuracy. There are six layers in the network, with each contributing uniquely towards adapting to varying features. The initial layer of NACN is a self-attention-based feature alignment module that maps the chosen feature subset into a shared latent space. Given an input feature set , where ≤ (selected features), the feature transformation is performed as follows: 30 where and are trainable parameters, and is the unified latent space dimension. A self-attention mechanism refines the feature representation: 31 where and help reweight the features, ensuring the network retains the most relevant information. Because the number of selected features d’ can change, we propose a dynamic fully connected (FC) layer that learns feature mappings for various input dimensions. The transformation is represented as: 32 This layer allows the classifier to adapt its learned representations depending on the input features dynamically. In order to standardize the output dimension prior to classification, we use an adaptive pooling layer that makes sure all feature representations are of equal length. The layer sums feature values via an averaging function: 33 This helps to maintain stability between various feature subsets while preserving critical information. To provide additional stability and avoid overfitting, a batch normalization layer is used to normalize and shift the pooled features: 34 where μ and σ are batch statistics, and ε is a small constant to avoid division by zero. This helps to ensure that the classifier does not degrade with variability in feature distributions. Because the classifier needs to generalize well across many feature subsets, we use meta-learning, in which the network is trained on different feature sets. The classification function is given by: 35 where and , with number of classes. We employ a combination of categorical cross-entropy loss and L2 regularization to prevent overfitting and optimize both the training and testing losses. The optimization step is given in Eq. ( 24 ). 36 where is the true label, is the predicted probability, and λ is the regularization parameter. Conclusion and future scope This study introduces a novel Content-Based Image Retrieval (CBIR) framework that combines a Casualty-Driven Multiresolution RESNET-50 architecture with a variational autoencoder and fast osprey optimization algorithm to reduce feature dimensions. Experiments were conducted on three benchmark datasets, namely CIFAR-10, Oxford Flower, and Corel 1000, employing the Structural Similarity Index and Histogram Intersection as similarity measures. The proposed CBIR framework demonstrated superior performance, achieving accuracies of 0.9795 and 0.9861, precisions of 0.9685 and 0.9856, recalls of 0.9705 and 0.9870, and F-scores of 0.9695 and 0.9895, respectively, across the CIFAR-10, Oxford Flowers, and Corel 1000 datasets. These results confirm the high efficiency of the proposed model, offer an accurate and robust image retrieval technique, and establish a new benchmark for CBIR systems. Future enhancements could involve integrating new techniques with CBIR to further improve its performance. Examples include employing transformers and vision-language models to enhance the understanding of contextual features and cross-modal retrieval processes. Lightweight CNN architecture 62 , Multimodal data retrieval 63 , and Transfer Learning 64 can be used in domain adaptation to address domain-specific retrieval tasks. Explainable AI (XAI) can be integrated into the framework to improve transparency and provide insights into model behavior under noisy and augmented data conditions 5 . Funding No funding. Data availability The datasets used in this study are publicly available benchmark datasets, including CIFAR-10, Oxford Flowers, and Corel-1000. These datasets are widely used for academic research and are accessible through their respective official repositories. All data used in this work are available from the corresponding sources and have been properly cited in the manuscript. The use of these datasets complies with their respective licensing terms for non-commercial research purposes. Further details regarding dataset access and usage can be provided by the corresponding author upon reasonable request. Code availability The code supporting the findings of this study is publicly available in a GitHub repository at: https://github.com/abdulrahmanzeain-wq/CBIR-Code-.git . The repository contains the full implementation of the proposed method along with the necessary scripts to reproduce the experimental results. An archived and citable version of the code has been deposited in Zenodo and is accessible via the following 10.5281/zenodo.19235286. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Jain, S., Pulaparthi, K. & Fulara, C. Content based image retrieval. Int. J. Adv. Eng. Glob. Technol. 3 (10), 1251–1258 (2015). [ Google Scholar ] 2. Abdullah, M. N., and Sahari, M. A. Digital image clustering and colour model selection in content-based image retrieval (CBIR) approach for biometric security images. In AIP Conference Proceedings , Vol. 2617, no. 1, 020006 (AIP Publishing LLC, 2022). 3. Chen, W. et al. Deep learning for instance retrieval: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 45 (6), 7270–7292 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Aggarwal, A. K. Explainable and uncertainty-aware deep learning for image classification under noisy and augmented conditions. Int. J. Signal Process. 10 , 39–42 (2025). [ Google Scholar ] 5. Aggarwal, A. K. A performance evaluation of deep learning-based image classification models on noisy and augmented datasets.. Int. J. Electr. Eng. Comput. Sci. 7 , 136–141 (2025). [ Google Scholar ] 6. Kumar, S. et al. A hierarchical approach based CBIR scheme using shape, texture, and color for accelerating retrieval process. J. King Saud Univ. Comput. Inf. Sci. 35 (7), 101609 (2023). [ Google Scholar ] 7. Liu, Y., Ding, L., Chen, C. & Liu, Y. Similarity-based unsupervised deep transfer learning for remote sensing image retrieval. IEEE Trans. Geosci. Remote Sens. 58 (11), 7872–7889 (2020). [ Google Scholar ] 8. Raju, G. K., Padmanabham, P., and Govardhan, A. Enhanced content-based image retrieval with trio-deep feature extractors with multi-similarity function. Int. J. Intell. Eng. Syst. 15 (6), 511–525 (2022). 9. Pascal, A., Planas, A., Vidal, Z. F., Bonti, A., Tonelotto, L., and Castiglioni, L. (2024). Image feature extraction for similarity searching using transfer learning with ResNet. In XXX Congreso Argentino de Ciencias de la Computación (CACIC)(La Plata, 7 al 11 de octubre de 2024) . 10. Roy, S., Sangineto, E., Demir, B. & Sebe, N. Metric-learning-based deep hashing network for content-based retrieval of remote sensing images. IEEE Geosci. Remote Sens. Lett. 18 (2), 226–230 (2020). [ Google Scholar ] 11. Qin, Q., Huang, L., Wei, Z., Xie, K. & Zhang, W. Unsupervised deep multi-similarity hashing with semantic structure for image retrieval. IEEE Trans. Circuits Syst. Video Technol. 31 (7), 2852–2865 (2020). [ Google Scholar ] 12. Li, X. et al. From matching to generation: A survey on generative information retrieval. ACM Trans. Inf. Syst. 43 (3), 1–62 (2025). [ Google Scholar ] 13. Zhang, H., Zhao, P., Miao, X., Shao, Y., Liu, Z., Yang, T., and Cui, B. (2023). Experimental analysis of large-scale learnable vector storage compression. arXiv preprint arXiv:2311.15578 . 14. Tong, X. Y. et al. Exploiting deep features for remote sensing image retrieval: A systematic investigation. IEEE Trans. Big Data 6 (3), 507–521 (2019). [ Google Scholar ] 15. Yue, H. et al. Deep learning for diagnosis and classification of obstructive sleep apnea: A nasal airflow-based multi-resolution residual network.. Nat. Sci. Sleep 10.2147/NSS.S297856 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Wang, X. et al. A review of image super-resolution approaches based on deep learning and applications in remote sensing. Remote Sensing 14 (21), 5423 (2022). [ Google Scholar ] 17. Dehghani, M. & Trojovský, P. Osprey optimization algorithm: A new bio-inspired metaheuristic algorithm for solving engineering optimization problems. Front. Mech. Eng. 8 , 1126450 (2023). [ Google Scholar ] 18. Giuste, F. O., and Vizcarra, J. C. Cifar-10 image classification using feature ensembles. arXiv preprint arXiv:2002.03846 (2020). 19. Nilsback, M. E., and Zisserman, A. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing 722–729 (IEEE, 2008). 20. Artemi, M., and Liu, H. Image optimization using improved gray-scale quantization for content-based image retrieval. In 2020 IEEE 6th International Conference on Optimization and Applications (ICOA) 1–6 (IEEE, 2020). 21. Mansour, E. & Alsaud, A. B. F. B. M. Advanced information retrieval techniques in the big data era: Trends, challenges, and applications. Metall. Mater. Eng. 31 (4), 466–483 (2025). [ Google Scholar ] 22. Sallam, M. & Shnan, M. A. Enhancing semantic image retrieval using self-supervised learning: A label-efficient approach. Babylon. J. Mach. Learn. 2025 , 42–60 (2025). [ Google Scholar ] 23. Luo, Y., Li, W., Ma, X. & Zhang, K. Image retrieval algorithm based on locality-sensitive hash using convolutional neural network and attention mechanism. Information 13 (10), 446 (2022). [ Google Scholar ] 24. Jeya Christy, A. & Dhanalakshmi, K. Content-based image recognition and tagging by deep learning methods. Wirel. Pers. Commun. 123 (1), 813–838 (2022). [ Google Scholar ] 25. Bibi, R., Mehmood, Z., Munshi, A., Yousaf, R. M. & Ahmed, S. S. Deep features optimization based on a transfer learning, genetic algorithm, and extreme learning machine for robust content-based image retrieval. PLoS ONE 17 (10), e0274764 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Rastegar, H., Giveki, D. & Choubin, M. EEG signals classification using a new radial basis function neural network and jellyfish meta-heuristic algorithm. Evol. Intell. 17 (2), 1197–1208 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Rastegar, H. & Giveki, D. Designing a new deep convolutional neural network for content-based image retrieval with relevance feedback. Comput. Electr. Eng. 106 , 108593 (2023). [ Google Scholar ] 28. Giveki, D., Soltanshahi, M. A. & Rastegar, H. Shape classification using a new shape descriptor and multi-view learning. Displays 82 , 102636 (2024). [ Google Scholar ] 29. Giveki, D. & Esfandyari, S. Semantic image representation for image recognition and retrieval using multilayer variational auto-encoder, InceptionNet and low-level image features. J. Supercomput. 81 (1), 346 (2025). [ Google Scholar ] 30. Giveki, D., Zaheri, A. & Allahyari, N. Designing CNNs with optimal architectures using antlion optimization for plant leaf recognition. Multimed. Tools Appl. 84 (8), 4625–4654 (2025). [ Google Scholar ] 31. Farsijani, F., Zaheri, A., Giveki, D., and Peyvandi, H. Low-level feature representation in the RCSU-Net model using channel-spatial attention mechanism for semantic segmentation of plant leaves. In 2025 29th International Computer Conference, Computer Society of Iran (CSICC) 1–5 (IEEE, 2025). 32. Singh, M. & Singh, M. K. Content-based gastric image retrieval using fusion of deep learning features with dimensionality reduction. SN Comput. Sci. 6 (2), 1–12 (2025). [ Google Scholar ] 33. Rani, S., Kasana, G. & Batra, S. An efficient content based image retrieval framework using separable CNNs. Clust. Comput. 28 (1), 56 (2025). [ Google Scholar ] 34. Mowery, B. D. The paired t-test.. Pediatr. Nurs. 37 (6), 320–322 (2011). [ PubMed ] [ Google Scholar ] 35. Krizhevsky, A. Learning Multiple Layers of Features from Tiny Images Technical Report, University of Toronto (2009). 36. Nilsback, N., and Zisserman, A. Automated flower classification over a large number of classes. In: Proceedings of the 7th Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP) (2008). 37. Tiwari, A. A hybrid feature selection method using an improved binary butterfly optimization algorithm and adaptive β–hill climbing. IEEE Access 11 , 93511–93537 (2023). [ Google Scholar ] 38. Tiwari, A. & Chaturvedi, A. Automatic channel selection using multiobjective X-shaped binary butterfly algorithm for motor imagery classification. Expert Syst. Appl. 206 , 117757 (2022). [ Google Scholar ] 39. Sujon, K. M., Hassan, R., Choi, K. & Samad, M. A. Accuracy, precision, recall, f1-score, or MCC? Empirical evidence from advanced statistics, ML, and XAI for evaluating business predictive models.. J. Big Data 12 (1), 268 (2025). [ Google Scholar ] 40. Miao, J. & Zhu, W. Precision–recall curve (PRC) classification trees. Evol. Intell. 15 (3), 1545–1569 (2022). [ Google Scholar ] 41. Wang, B. (2022). A parallel implementation of computing mean average precision. arXiv preprint arXiv:2206.09504 . 42. Dupret, G. Discounted cumulative gain and user decision models. In International Symposium on String Processing and Information Retrieval 2–13 (Springer, Berlin, 2011). 43. Latha, D. & Jacob Vetha Raj, Y. Hybrid CBIR method using statistical, DWT‐entropy and POPMV‐based feature sets. IET Image Process. 13 (12), 2031–2044 (2019). [ Google Scholar ] 44. Ahmed, K. T., Ummesafi, S. & Iqbal, A. Content based image retrieval using image features information fusion.. Information Fusion 51 , 76–99 (2019). [ Google Scholar ] 45. Alsmadi, M. K. Content-based image retrieval using color, shape and texture descriptors and features.. Arab. J. Sci. Eng. 45 (4), 3317–3330 (2020). [ Google Scholar ] 46. Khan, U. A., Javed, A. & Ashraf, R. An effective hybrid framework for content based image retrieval (CBIR). Multimed. Tools Appl. 80 (17), 26911–26937 (2021). [ Google Scholar ] 47. Ashraf, R. et al. MDCBIR-MF: Multimedia data for content-based image retrieval by using multiple features. Multimed. Tools Appl. 79 (13), 8553–8579 (2020). [ Google Scholar ] 48. Keisham, N., and Neelima, A. Content-based medical image retrieval using CNN feature extraction and hashing. In Deep Learning in Biomedical Signal and Medical Imaging 179–188 (CRC Press, 2025). 49. Anju, J. & Shreelekshmi, R. A faster secure content-based image retrieval using clustering for cloud. Expert Syst. Appl. 189 , 116070 (2022). [ Google Scholar ] 50. Ma, W. et al. A privacy-preserving content-based image retrieval method based on deep learning in cloud computing. Expert Syst. Appl. 203 , 117508 (2022). [ Google Scholar ] 51. Zhang, K. et al. Content-based image retrieval with a Convolutional Siamese Neural Network: Distinguishing lung cancer and tuberculosis in CT images. Comput. Biol. Med. 140 , 105096 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 52. Taheri, F., Rahbar, K. & Salimi, P. Effective features in content-based image retrieval from a combination of low-level features and deep Boltzmann machine. Multimed. Tools Appl. 82 (24), 37959–37982 (2023). [ Google Scholar ] 53. Wang, Y., Chen, L., Wu, G., Yu, K. & Lu, T. Efficient and secure content-based image retrieval with deep neural networks in the mobile cloud computing. Comput. Secur. 128 , 103163 (2023). [ Google Scholar ] 54. Sudhish, D. K. & Nair, L. R. Content-based image retrieval for medical diagnosis using fuzzy clustering and deep learning. Biomed. Signal Process. Control 88 , 105620 (2024). [ Google Scholar ] 55. Alrahhal, M. & Supreethi, K. P. Integrating machine learning algorithms for robust content-based image retrieval. Int. J. Inf. Technol. 16 (8), 5005–5021 (2024). [ Google Scholar ] 56. Woolson, R. F. Wilcoxon signed‐rank test. Wiley encyclopedia of clinical trials , 1–3 (2007). 57. Wang, Z., Bovik, A. C., and Sheikh, H. R. Structural similarity based image quality assessment. In Digital Video Image Quality and Perceptual Coding 225–242 (CRC Press, 2017). 58. Barla, A., Odone, F., and Verri, A. Histogram intersection kernel for image classification. In Proceedings 2003 international conference on image processing (Cat. No. 03CH37429) , Vol. 3, III-513 (IEEE, 2003). 59. Musa, P., Al Rafi, F., and Lamsani, M. A Review: Contrast-Limited Adaptive Histogram Equalization (CLAHE) methods to help the application of face recognition. In 2018 Third International Conference on Informatics and Computing (ICIC) 1–6 (IEEE, 2018). 60. Gal, M. S. & Rubinfeld, D. L. Data standardization. NYUL Rev. 94 , 737 (2019). [ Google Scholar ] 61. Deatrick, J. A., Knafl, K. A. & Murphy‐Moore, C. Clarifying the concept of normalization. Image 31 (3), 209–214 (1999). [ DOI ] [ PubMed ] [ Google Scholar ] 62. Thukral, R., Arora, A. S., Aggarwal, A. K., Dora, T., and Sancheti, S. Effectiveness of passive thermography in predicting mucositis in head and neck cancer patients using light-weight convolutional neural networks approach. In International Conference on Frontiers of Intelligent Computing: Theory and Applications 1–13 (Springer Nature Singapore, Singapore, 2024). 63. Yan, J. et al. Research on multimodal techniques for arc detection in railway systems with limited data. Struct. Health Monit. 10.1177/147592172513367 (2025). [ Google Scholar ] 64. Jin, B., Cruz, L. & Gonçalves, N. Deep facial diagnosis: Deep transfer learning from face recognition to facial diagnosis. IEEE Access 8 , 123649–123661 (2020). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The datasets used in this study are publicly available benchmark datasets, including CIFAR-10, Oxford Flowers, and Corel-1000. These datasets are widely used for academic research and are accessible through their respective official repositories. All data used in this work are available from the corresponding sources and have been properly cited in the manuscript. The use of these datasets complies with their respective licensing terms for non-commercial research purposes. Further details regarding dataset access and usage can be provided by the corresponding author upon reasonable request. The code supporting the findings of this study is publicly available in a GitHub repository at: https://github.com/abdulrahmanzeain-wq/CBIR-Code-.git . The repository contains the full implementation of the proposed method along with the necessary scripts to reproduce the experimental results. An archived and citable version of the code has been deposited in Zenodo and is accessible via the following 10.5281/zenodo.19235286. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (6.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Learning to Restore More: Continual Capability Expansion for Pretrained Image Restoration Models#799517
• Learning to Restore More: Continual Capability Expansion for Pretrained Image Restoration Models#798094
• Boosting lightweight CNN-based networks via selective residual attentive patterns for image recognition#414969
• Causal network structure predicts memory organization and neural reinstatement across events#478008
• BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models#16564