1
Causal Disentanglement for Full-Reference Image Quality Assessment
arXiv:2604.21654v1 [cs.CV] 23 Apr 2026
Zhen Zhang, Jielei Chu, Senior Member, IEEE, Tian Zhang, Weide Liu, Fengmao Lv, Tianrui Li, Senior Member, IEEE, Jun Cheng, Senior Member, IEEE, Yuming Fang, Fellow, IEEE,
Abstract—Existing deep network-based full-reference image quality assessment (FR-IQA) models typically work by performing pairwise comparisons of deep features from the reference and distorted images. In this paper, we approach this problem from a different perspective and propose a novel FR-IQA paradigm based on causal inference and decoupled representation learning. Unlike typical feature comparisonbased FR-IQA models, our approach formulates degradation estimation as a causal disentanglement process guided by intervention on latent representations. We first decouple degradation and content representations by exploiting the content invariance between the reference and distorted images. Second, inspired by the human visual masking effect, we design a masking module to model the causal relationship between image content and degradation features, thereby extracting content-influenced degradation features from distorted images. Finally, quality scores are predicted from these degradation features using either supervised regression or label-free dimensionality reduction. Extensive experiments demonstrate that our method achieves highly competitive performance on standard IQA benchmarks across fully supervised, few-label, and label-free settings. Furthermore, we evaluate the approach on diverse non-standard natural image domains with scarce data, including underwater, radiographic, medical, neutron, and screen-content images. Benefiting from its ability to perform scenario-specific training and prediction without labeled IQA data, our method exhibits superior cross-domain generalization compared to existing training-free FR-IQA models. Index Terms—Image quality assessment, FR-IQA, label-free, visual masking effect.
I. Introduction Zhen Zhang is with the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China. (e-mail: [email protected]). Tian Zhang is with the School of Physics, Northeast Normal University, Changchun 130024, China (e-mail: [email protected]). Jun Cheng is with the Institute for Infocomm Research, Agency for Science, Technology and Research, Singapore 138632. (e-mail: [email protected]). Weide Liu is with School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics, Nanchang, 330013, China. ([email protected]). Yuming Fang is with the School of Information Management, Jiangxi University of Finance and Economics, Nanchang, China. (email: [email protected]). Jielei Chu, Fengmao Lv, Tianrui Li are with the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China. Jielei Chu is also with the Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education, Chengdu, Sichuan 611756, China. Tianrui Li is also with National Engineering Laboratory of Integrated Transportation Big Data Application Technology, Southwest Jiaotong University, Chengdu 611756, China. (e-mail: {fengmaolv, jieleichu, trli}@swjtu.edu.cn). Jielei Chu is the corresponding author.
Fig. 1. (a) shows the typical FR-IQA architecture. Most deep FRIQA models compute the differences between feature maps to predict quality scores. (b) shows the proposed FR-IQA architecture, where we extract content-influenced degradation features to predict quality scores.
I
MAGE quality assessment can help operators screen and provide high-quality digital images to their subscribers. Therefore, developing image quality assessment (IQA) methods that align well with human subjective perception becomes extremely important. IQA can be categorized into full-reference IQA (FR-IQA) [1]–[3], reducedreference IQA (RR-IQA) [4], and no-reference IQA (NRIQA) [5]. FR-IQA models are generally believed to be less practical than NR-IQA models because of their reliance on pristine images. However, recent studies [6]–[8] have shown that FR-IQA models can guide perceptual image enhancement, a task that poses challenges for many NRIQA models [9]. As such, FR-IQA models remain crucial. FR-IQA obtains quality assessment results by determining the difference between the distorted image and the reference image. In previous studies, this difference was usually regarded as a measure of the distance between these images. As shown in Figure 1, deep network-based FR-IQA models often employ deep learning for feature decomposition and score pooling. These FR-IQA models typically extract feature maps from both distorted and reference images, and compute their differences to derive quality scores. Deep network-based FR-IQA methods can be categorized into end-to-end deep learning models (training-dependent) and training-free models. Although these methods have achieved great success in the field of FR-IQA, several limitations remain. Training-dependent methods [6], [10], [11] rely on large-scale labeled IQA data. However, due to the high subjectivity of human
2
perception, obtaining reliable ground truth typically necessitates a large panel of human observers to conduct exhaustive scoring or ranking (e.g., Mean Opinion Scores). Consequently, this heavy reliance restricts their scalability and applicability, particularly in data-scarce or emerging domains (e.g., infrared or underwater images). Moreover, training-free [12]–[15] methods typically leverage pretrained models (e.g., VGG [16], EfficientNet [17]) trained on ImageNet to extract feature maps from both reference and distorted images. They then compute the distance between these maps to derive the quality scores. However, these methods cannot be readily adapted to specific datasets, leading to severe performance drops in some scenarios. For instance, DBIQA [11] demonstrated that DeepWSD [18] only achieves excellent performance in limited scenarios. Therefore, it is essential to develop an FR-IQA method that does not rely on subjective quality annotations and can be efficiently adapted to specific datasets. In this work, we propose a method based on causal inference and decoupled representation learning that can be efficiently adapted to diverse scenarios (e.g., image enhancement or radiographic imaging) without the need for costly Mean Opinion Score (MOS) annotations. To achieve this label-free adaptability, the key lies in establishing unsupervised paradigms for both training and prediction. For the training phase, our approach is built upon the premise that a distorted image fundamentally comprises two components: image content and image degradation. In FR-IQA, distorted images are typically synthesized by applying degradations to reference images. This generation process not only validates this premise but also guarantees that the image content remains strictly identical between the reference and the distorted images. By exploiting this image content invariance, we can decouple the degradation features from the image content in an unsupervised manner. However, merely decoupling image degradation is insufficient to fully simulate human subjective perception. According to the visual masking effect (VME), the Human Visual System (HVS) exhibits varying sensitivities to the same distortion depending on the underlying image content. For example, noise is highly visible in smooth areas (e.g., a clear sky) but nearly imperceptible in textured ones (e.g., dense grass). This phenomenon reveals that human perception of image degradation is affected by the interaction between image content and distortion. Because image content influences the visibility of degradation, it is imperative to explicitly model how image content influences the degradation representation. To this end, we propose a method based on causal inference and decoupled representation learning to extract content-influenced degradation features. In the prediction phase, employing supervised or few-shot regression allows the model to explicitly map these causal features to subjective scores, thereby achieving a finer alignment with human visual perception. When MOS annotations are unavailable, absolute quality prediction is not identifiable without additional supervision. We therefore formulate zero-shot quality prediction as a man-
ifold learning problem. Specifically, based on the analysis above, the learned content-influenced degradation feature is assumed to be primarily governed by a latent perceptual quality variable, so that its quality-relevant variation lies near a low-dimensional manifold in the high-dimensional feature space. Under this view, the intrinsic geometry of the feature manifold reflects the relative perceptual severity of distortions. We thus employ UMAP to recover a one-dimensional embedding that preserves the local neighborhood structure of the learned feature space. The resulting coordinate serves as a relative quality descriptor, enabling zero-shot quality ranking and discrimination without relying on MOS supervision. The main contributions of this work are summarized as follows. • We propose a new FR-IQA paradigm based on decoupled representation learning, departing from traditional double-branch approaches. The paradigm uses do-operator to guide latent representations generation, effectively extracting degradation features through content invariance between image pairs. • Inspired by the Visual Masking Effect (VME), we analyze the causal path of image content, degradation features, and perceived quality, introducing structural causal priors within the framework of decoupled representation learning. The causal layer can model the causal effect of image content on degradation. • Extensive experiments demonstrate that the proposed method achieves highly competitive performance on standard IQA benchmarks. Our model exhibits exceptional robustness across fully supervised, few-label, and completely label-free settings. • Comprehensive cross-domain experiments on diverse, non-standard image domains (e.g., underwater, radiographic, and medical images) validate the method’s superior generalization capability. Its ability to perform adaptation without labeled IQA data highlights its immense application potential in data-scarce domains. This paper is structured as follows. In Section II, we review the existing work related to our study. In Section III-A, we present the problem formulation of the proposed FR-IQA framework. In Section IV, we describe the proposed method in detail. In Section V, we report the model settings and experimental results. In Section VI, we discuss the limitations and failure cases of the proposed method. Finally, the conclusions are presented in Section VII. II. Related Work A. Various Quality Assessment Scenarios The development of FR-IQA models has never been detached from practical quality assessment scenarios. Early benchmark datasets, such as TID2013, LIVE, CSIQ, and KADID-10k, were mainly developed to study human perception of fidelity loss caused by synthetic distortions under controlled settings. More recent datasets, including
3
PIPAL, QADS, SynTEX, and TQD, further extend FRIQA to images generated by modern restoration and synthesis algorithms, where perceptual quality is often harder to determine because algorithm-induced changes do not always correspond to quality degradation. Meanwhile, FR-IQA has also been explored in more specific scenarios, such as screen-content, infrared, medical, and neutron images. However, progress in these areas remains limited, largely because IQA annotation is costly, subjective, and often requires domain expertise, making labeled datasets difficult to obtain. Moreover, existing training-free FRIQA methods show performance degradation in crossdomain scenarios, since they usually rely on ImageNetpretrained feature representations that are primarily optimized for natural images. When applied to domain-specific data, the discrepancy in image statistics and structural information leads to domain shift and compromises prediction reliability. These limitations call for an FR-IQA framework that can be adapted to domain-specific features without relying on subjective quality labels. B. Deep Network-Based FR-IQA Models Deep networks have been widely used to model the complex relationship between reference and distorted images in FR-IQA. Early studies extracted sensitivity or similarity maps from both images for quality estimation [19], [20]. Later, researchers found that VGG-based deep features effectively represent perceptual similarity, serving as useful losses for image synthesis and quality prediction [21]. Building on this, Ding et al. [6] proposed DISTS, which measures structural and textural similarity in deep feature space, and ADISTS [22] further improved it using a dispersion index for adaptive weighting. Liao et al. [18] instead modeled perceptual degradation using Wasserstein distance in feature space, emphasizing distributional differences for quality evaluation. TOPIQ [10] introduced a top-down approach featuring a cross-scale attention mechanism, which leverages high-level semantics to guide the network’s focus toward important local distortion regions for efficient quality evaluation. DBIQA [11] proposed a dual-branch framework using kernel representation similarity analysis (KRSA) and mean absolute error (MAE). This method focuses on correlations among features from the same source. DeepCausal [13] proposed an FR-IQA method using abductive counterfactual inference to uncover causal links between deep network features and perceptual distortions, thereby achieving a trainingfree FR-IQA framework. DeepDC [15] utilized distance correlation to quantify statistical similarity in deep feature space, thereby overcoming the strict requirement of pixellevel alignment for image quality evaluation. DeepJSD investigated various deep distribution measures, such as Jensen-Shannon divergence and symmetric KullbackLeibler divergence, to perceptually compare pre-trained network features for a training-free FR-IQA framework. However, existing methods still face practical constraints: training-dependent models are limited by the
scarcity of annotated IQA data, while training-free models struggle to adapt quickly to domain-specific datasets. Despite these bottlenecks, these methods have achieved considerable success by formulating FR-IQA as a measure of deep feature differences. In this paper, we propose a novel perspective: formulating FR-IQA as a causal disentanglement representation problem grounded in the visual masking effect. C. Causal Disentanglement Representation Learning Disentangled representation learning aims to separate the underlying generative factors within data into independent and interpretable components. Early approaches, such as Independent Component Analysis (ICA) [23], attempted to decompose signals into statistically independent sources. Variational Autoencoders (VAE) [24] further advanced this idea by introducing regularization and prior constraints to structure the latent space, while β-VAE [25] enhanced disentanglement through KL divergence optimization. However, traditional methods often assume independent latent factors, which is unrealistic for data with complex causal dependencies. To address this, CausalVAE [26] incorporates causal structures to achieve causal disentanglement, and SCADI [27] employs pseudo-label generation for semi-supervised causal disentanglement. In this paper, we formulate FR-IQA from the perspective of causal disentanglement representation learning. D. Visual Masking Effects Recent research reveals that image content exerts a complex causal influence on the visibility of distortions during visual perception [28]–[31]. The visual quality of an image mainly depends on four factors: contrast sensitivity function (CSF), luminance adaptation (LA), contrast masking (CM), and foveated masking (FM). Spatial CSF indicates the bandpass properties of the HVS in the spatial frequency domain [28]. Compared with high-frequency regions, the HVS is more sensitive to distortion in low-frequency regions. LA reflects HVS sensitivity as influenced by background luminance [29]. It is harder to distinguish distortion in relatively dark or bright environments. CM refers to the attenuation of one contrast in the presence of another [30]. FM implies that visual sensitivity is related to retinal eccentricity, with distortion far from the focus of gaze being less perceptible [31]. In IQA, visual masking is commonly modeled to mimic human subjective perception [32], [33]. III. Preliminaries A. Problem Formulation As shown in Figure 2 (a), we construct a Structural Causal Model (SCM) for distorted image generation to describe distorted image generation process. In this SCM, C and D respectively denote the image content and degradation. The path C → Id ← D indicates that the
4
Fig. 2. Structural Causal Model (SCM) for FR-IQA. Left: distorted image generation mechanism; Right: visual mechanism after decoupling.
distorted image Id is generated from both image content C and degradation D. This SCM is based on the following proposition: Proposition 1. A distorted image Id can be predominantly described by two factors: (1) image content C, which represents the semantic and structural elements, and (2) degradation D, which represents all distortions and quality degradation affecting the image. Proposition 1 a reasonable formulation within the FRIQA framework. By definition, the reference image Ir is distortion-free, thereby providing the pristine image content C. Accordingly, the distorted image Id is formulated as a combination of this image content, which strictly corresponds to Ir , and an externally applied degradation D. This process aligns with the causal generation path C → Id ← D. Building upon this generation mechanism and Proposition 1, a question arises: Can we reformulate the FRIQA task as a decoupling process? FR-IQA assesses image quality by comparing a distorted image with its reference, with the only difference being the introduced degradation. An intuitive idea is to decouple the degradation D from Id using the pristine content C provided by Ir , and predict the Mean Opinion Score (MOS) based on the degradation D. However, this straightforward decoupling presents a critical limitation: relying solely on the isolated degradation D fails to accurately reflect human visual perception, because image content inherently masks or accentuates distortions due to the visual masking effect. The principle of visual masking indicates that image quality is mainly related to the visibility of distortion. Specifically, the same image shows different visual quality when degraded by distortions of different types and intensity levels. However, the visibility of an identical distortion varies across different image contents. As shown in Figure 3 (a), for lens blur, images with smooth content have higher MOS values than those with complex textures under the same distortion level, indicating that smooth content has a stronger masking ability for this distortion. However, images with smooth content are more sensitive to quantization noise than those with complex textures (see Figure 3 (b)). Therefore, although the degradation D acts as the fundamental determinant of quality assessment, it is modulated by the image content C due to the visual masking
Fig. 3. Illustrations of different distortion visibility. Under the same distortion intensity level, the image with smooth content can be more tolerant to lens blur than the image with textured content, yet more sensitive to quantization noise. The value in parentheses indicates the imposed distortion intensity level.
effect. Motivated by this insight, we reformulate FRIQA as a process of causal disentanglement representation learning, as shown in Figure 2 (b). First, we decouple degradation features from Id under the guidance of content invariance between the reference and distorted images. Second, we model the causal influence of the image content on these degradation features, C → D. Finally, the resulting content-influenced degradation feature is used to predict the MOS. B. Causal Disentanglement Representation Learning Causal disentanglement representation learning seeks to learn latent variables that are semantically interpretable and structured according to an underlying causal generative process. In contrast to conventional disentangled representation learning, which typically assumes statistical independence among latent factors, causal disentanglement explicitly models structural dependencies by introducing a causal graph over the latent variables. Under this formulation, the observed data are assumed to be generated from independent exogenous variables, which are subsequently transformed into causally dependent endogenous representations [26], [27].
5
Encoder
Decoder
Clear image
Fine-tuning on the corresponding dataset
Synthetic degeneration
Synthetic distorted image
Pre-training
Degenerate features
Reconstructed image
Distorted image
Causal effect Reference image
Fine-tuning
Reconstructed image
Fig. 4. In the pre-training stage, we first construct a synthetic degraded image dataset from clear images. Subsequently, the model is pre-trained using the constructed synthetic degradation dataset. In the fine-tuning stage, we fine-tune the model on specific IQA datasets (e.g., TID2013 [34], LIVE [35]) or other datasets containing distorted images and reference images. The fine-tuning model remains identical to those used in pre-training, with the only difference being the dataset.
Let ϵ ∈ Rn denote the independent exogenous variables, and let z ∈ Rn denote the endogenous latent representations endowed with causal semantics. Given an adjacency matrix A that encodes the causal relations among latent variables, the causal layer is defined as z = A⊤ z + ϵ = (I − A⊤ )−1 ϵ,
ϵ ∼ N (0, I),
(1)
where A specifies the directed dependencies among the latent variables. This formulation indicates that the latent representation z is no longer a collection of mutually independent factors, but instead a structured representation produced through causal propagation over the graph. To ensure that the learned causal structure is welldefined, the adjacency matrix A must correspond to a directed acyclic graph (DAG). CausalVAE imposes the following continuous acyclicity constraint: )n ) (( c − n = 0, (2) H(A) ≡ tr I + A ◦ A m where tr(·) denotes the trace of a matrix, m is a positive scaling factor, ◦ denotes element-wise multiplication, c is a positive constant, and n is the number of latent variables. This constraint guarantees that A represents a valid DAG and thereby ensures that the learned latent variables admit a causal interpretation. Accordingly, causal disentanglement representation learning can be understood as the process of learning semantically meaningful latent factors together with their causal relationships, rather than merely enforcing statistical independence in the latent space [27], [36]. IV. Method The whole framework consists of three stages. Figure 4 shows the process of pre-training and fine-tuning. In the pre-training stage, we first construct a synthetic degraded image dataset from clear images. Subsequently, the model is pre-trained using the constructed synthetic degradation dataset. In the fine-tuning stage, we fine-tune the model on specific IQA datasets (e.g., TID2013 [34], LIVE [35]) or other datasets containing distorted images and reference
images. The fine-tuning model remains identical to the model used in pre-training, with the only difference being the dataset. In the zero-shot setting, the framework proceeds solely with pre-training, without subsequent fine-tuning. Finally, in the prediction stage, we provide multiple approaches (i.e., supervised, few-shot, and zeroshot) to predict the quality score. Sections IV-A and IV-B describe the training methods for the pre-training and fine-tuning stages. Section IV-C explains how to predict the quality score. Section V presents the details of the proposed method. A. Decouple degradation features Guided by the problem formulation in Section III-A, our objective at this stage is to separate distortion-related factors from content-related factors in distorted images. A naive approach is to use a standard autoencoder without intervention. We use an SCM to describe the latent generation process without intervention, as illustrated in Figure 5 (a). The path Id → Z → Iˆd indicates that Z is the latent representation learned from the distorted image Id . Even if the autoencoder reconstructs the distorted image accurately, i.e., Iˆd ≈ Id , the latent representation may still remain entangled. Extracting decoupled degradation features without supervision is theoretically impossible for standard autoencoders; it fundamentally requires additional inductive biases or explicit interventions [37]. To support this separation, we exploit a task-specific property of FR-IQA: the reference image provides content information that is aligned with the distorted image. This property is stated as the following proposition. Proposition 2. Under the standard FR-IQA setting considered in this work, the reference image Ir and the distorted image Id share the same underlying scene content, while Id additionally contains distortion-related factors. Based on Proposition 1 and Proposition 2, we exploit the content invariance between the reference and distorted images to introduce an explicit intervention on the latent representation, thereby providing an inductive bias for
6
Id
D: Degradation features C: Reference images Id Id: Distorted images m: Mos
C
D
C
Id
m
Id
Causal
D
m
Fig. 5. (a) represents the structure of the autoencoder, and (b) is the structure of the autoencoder after intervention. Fig. 6. Structural causal model for predicting quality scores.
decoupling degradation features. As shown in Figure 5(b), let Z = (Z1:k , Zk+1:n ), (3) where Z1:k and Zk+1:n denote two complementary parts of the latent representation. Our intervention process is defined as follows. 1) Given a distorted image Id , the encoder E extracts its latent representation Z = (Z1:k , Zk+1:n ). 2) Let C denote the content information from the reference image Ir . We replace the first part Z1:k with C, yielding the intervened latent representation (C, Zk+1:n ). 3) We feed the intervened latent representation (C, Zk+1:n ) into the decoder de to generate the distorted image I˙d (C, Zk+1:n ). After intervention, the decoder reconstructs the distorted image from the image content and the remaining latent representation: I˙d = de(C, Zk+1:n ),
(4)
where C is the content from the reference image Ir . The corresponding reconstruction objective is ( ) d I˙d , Id = d(de(C, Zk+1:n ), Id ) , (5) where d(·, ·) denotes a reconstruction metric. If the decoder can reconstruct the distorted image well, namely de(C, Zk+1:n ) ≈ Id , Zk+1:n is encouraged to retain the degradation-related information. Since C is provided by the reference image and supplies the complete image content information for reconstruction, the complementary information required to reproduce the distortion must ( ) ˙ be retained in Zk+1:n . Therefore, minimizing d Id , Id in Eq. (5) encourages Zk+1:n to capture distortion-related factors. In this way, the intervention-reconstruction mechanism induces a functional separation in the latent space, where Zk+1:n gradually becomes a degradation-related representation. This provides the theoretical motivation for our intervention-based pre-training strategy. B. Causal representation learning While the framework established in Section IV-A can decouple degradation representations, relying solely on this degradation D is insufficient for accurately simulating
human visual perception. To formalize this discrepancy, we model the decoupling process with an SCM in Figure VII. From a causal perspective, the path C → I˙d ← D forms a causal collider [], implying that the ideal content C and the degradation D are independent [38], [39]. As previously analyzed in Section III-A, image content inherently masks or accentuates distortions due to the visual masking effect. Therefore, we model this masking effect by leveraging the ideal image content C and the decoupled degradation D. To model the content-influenced effect, we follow the causal disentanglement framework DEAR [36], which incorporates a causal model into the latent prior. In causal disentanglement representation learning, a set of latent variables z ∈ Rn is learned from observed data, while their dependencies are characterized by a DAG with weighted adjacency matrix A [23], [26], [27], [36], [40]. Let ϵ ∈ Rn denote the mutually independent exogenous variables, with ϵ ∼ N (0, I). Under the linear structural causal model, the causal layer is defined as z = A⊤ z + ϵ,
(6)
z = (I − A⊤ )−1 ϵ,
(7)
which yields where I is the identity matrix. To further enhance the flexibility of the causal transformation, we introduce element-wise nonlinear mappings and extend the structural equation as ( ) z = f (I − A⊤ )−1 h(ϵ) , (8) where f and h are element-wise nonlinear transformations. In this way, the independent exogenous variables ϵ are transformed into causally dependent latent representations z through nonlinear projection and graph-based causal propagation. In our problem formulation, the latent representation space is explicitly partitioned into only two components: the image content C and the decoupled degradation D. Different from generic causal disentanglement settings that infer an unknown latent causal graph, we explicitly impose the directed prior C → D according to the visual masking effect, namely, the visibility of degradation is
7
modulated by image content. Under this assumption, the adjacency matrix A degenerates into a two-node graph with a single directed edge from C to D. Consequently, the general nonlinear causal transformation in Eq. (8) reduces to a structural equation defined on the degradation node, yielding the content-conditioned degradation representation: Ḋ = V (h(C), D). (9) Here, V (h(C), D) models how semantic content modulates perceived degradations, allowing the network to capture content-influenced distortion perception. where h(C) denotes the content-related latent representation extracted from the reference image, and V (·) is a modulation function that combines content features and degradation features to generate the content-influenced degradation features Ḋ. As shown in Figure VII, intervention step 3) is modified so that the final distorted image is jointly generated by image contents C and Ḋ. The content-influenced degradation features Ḋ are used for subsequent prediction of the perceived quality. The detailed model design, including the encoder, causal layer, and decoder, is presented in Section V-B. C. Score prediction The final objective is to predict the image quality score based on the learned content-influenced degradation features Ḋ. Depending on the availability of labeled IQA data, we introduce two complementary prediction strategies. Supervised regression: When labeled IQA data are available, we employ a two-layer fully connected (FC) regression head to map the content-influenced degradation features Ḋ to the final quality scores. During this phase, all parameters of the model are jointly updated in an endto-end manner. This supervised fine-tuning further aligns the learned content-influenced degradation representation with subjective human perception. Dimensionality reduction: In the zero-shot setting where labeled IQA data are unavailable, absolute MOS values are generally not identifiable without additional supervision. Therefore, instead of directly regressing MOS, we aim to recover a relative quality coordinate from the contentinfluenced degradation feature D̃. We begin with the following assumption. Assumption 3. There exists a scalar latent perceptual quality variable q ∈ R such that q = v(C, D),
(10)
where v(·) denotes the content-dependent perceptual effect of degradation under visual masking. The subjective quality score is generated from q. m = ψ(q) + ϵ,
(11)
where ψ(·) is a monotone function and ϵ denotes subjective scoring noise. Accordingly, E[m | q] = ψ(q).
(12)
Assumption 3 is consistent with the problem formulation in Section III-A. Specifically, perceptual image quality is determined not by the physical degradation alone, but by its perceptual visibility after modulation by image content. Since human quality judgment is ultimately expressed as a scalar score, it is natural to model the underlying perceptual effect through a onedimensional latent variable q. Let D̃ ∈ Rp denote the ideal content-influenced degradation feature. We further assume that its dominant variation is governed by the latent quality variable q. Assumption 4. There exists an interval Q ⊂ R such that D̃ = Γ(q) + η,
q ∈ Q,
(13)
where Γ : Q → R is a mapping function and η is a small perturbation term. Moreover, there exist constants c1 , c2 > 0 such that for any q1 , q2 ∈ Q, p
c1 |q1 − q2 | ≤ ∥Γ(q1 ) − Γ(q2 )∥2 ≤ c2 |q1 − q2 |.
(14)
Equivalently, the quality-relevant features lie near a onedimensional manifold M = {Γ(q) : q ∈ Q} ⊂ Rp ,
(15)
and distances on this manifold are locally equivalent to distances in the latent quality variable. Assumption 4 states that although D̃ is embedded in a high-dimensional feature space, its quality-relevant geometry is intrinsically close to one-dimensional. This motivates the use of one-dimensional projection for relative quality prediction. Based on the above assumptions, we employ UMAP to project D̃ into a one-dimensional space. UMAP is used here as a practical embedding tool for recovering a one-dimensional coordinate from the learned feature geometry. In the following analysis, we only require that the resulting embedding approximately preserves the local neighborhood order and relative distance induced by D̃. Let yi ∈ R denote the one-dimensional embedding of the ith sample. Ideally, if sample i is closer to sample j than to sample k in the learned feature space, then the embedding tends to satisfy |yi − yj | < |yi − yk |.
(16)
The following proposition provides a justification for using the projected one-dimensional coordinate as a relative quality descriptor. Proposition 5. Suppose Assumptions 3 and 4 hold. The obtained one-dimensional embedding y approximately preserves the local neighborhood order and relative distance induced by D̃ on the sample support. Then there exists a monotone function T : Q → R such that y ≈ T (q),
(17)
up to perturbations caused by η and the embedding approximation error. Consequently, y can be used as a relative quality coordinate for ranking distorted images.
8
Relative distance
Direction Feature vector
Feature vector
Feature vector
Manifold approximation
Topological representation
Low-dimensional projection
Fig. 8.
Fig. 7. Local neighborhood relations in the feature space mainly reflect the proximity of the latent perceptual quality values.
In particular, the ordering induced by y is consistent with the ordering of the latent perceptual quality q, and hence with the ordering of E[m | q] = ψ(q). Proof. By Assumption 4, the ideal feature Γ(q) lies on a one-dimensional manifold parameterized by q ∈ Q. The bi-Lipschitz condition implies that, on Q, distances in the feature space are locally equivalent to distances in the latent variable. Therefore, when the perturbation term η is small, local neighborhood relations in D̃ remain consistent with those induced by q. Now consider a one-dimensional embedding y that approximately preserves the local neighborhood order of D̃ on the sample support. Since Q is an interval and the underlying structure is one-dimensional, preserving local neighborhood order implies that the recovered coordinate y provides an order-preserving parameterization of the latent variable. Hence there exists a monotone function T (·) such that y ≈ T (q), (18) where the approximation error is induced by the perturbation η and the imperfect preservation of neighborhood order during embedding. Finally, Assumption 3 gives E[m | q] = ψ(q),
(19)
where ψ(·) is monotone. Therefore, the coordinate y preserves the relative ordering of the latent perceptual quality and, equivalently, the ordering of the conditional expected subjective score. It thus serves as a valid relative quality coordinate for ranking distorted images. This result also clarifies the limitation of zero-shot prediction. Since the recovered coordinate is identifiable only up to an unknown monotone transformation, the onedimensional embedding can provide relative quality scores, but not an absolute MOS calibration. Therefore, absolute score prediction requires either a small number of labeled anchor points, as in QCN [41], or a lightweight regression head trained with few-shot IQA supervision. Specifically, given unseen distorted images Id and their reference images Ir , we feed them into the trained encoder and causal layer with frozen parameters to extract the
A simplified flowchart of the zero-shot prediction stage.
content-influenced degradation feature Ḋ. Then, UMAP is applied to project Ḋ into a one-dimensional space, yielding the relative quality coordinate y. Figure 8 shows the simplified flowchart of UMAP. Although this projection does not recover absolute MOS values, it preserves the relative ordering and local geometric structure of the learned quality manifold, which is sufficient for zero-shot quality ranking and discrimination. 6 5 4 3 2
Scores
Random initialization
6 5 4 3 2
Scores
After Pre-training
Fig. 9. Visualization of the one-dimensional UMAP embedding on TID2013. The color of each point becomes darker as the ground-truth quality decreases. After pre-training, the samples are continuously distributed along the one-dimensional space according to quality variation, while random initialization fails to produce a discriminative structure.
Figure 9 provides empirical evidence consistent with the above hypothesis. With random initialization, the projected points are densely mixed in the one-dimensional space, indicating that the learned features do not yet exhibit a meaningful quality manifold. After pre-training, however, the projected samples become continuously distributed according to quality variation, and high-quality samples can be clearly separated from low-quality ones. This observation suggests that pre-training enables the encoder and causal layer to organize the learned feature Ḋ around a one-dimensional quality-related manifold, making zero-shot ranking possible through UMAP projection. D. Connection to the HVS and Existing Works Connections to HVS: The proposed causal disentanglemen framework closely related to the HVS. In human visual perception, the visibility of a distortion is not absolute; it is heavily modulated by the spatial background, a psychophysical phenomenon known as the visual masking effect [28]–[31]. For instance, human eyes are significantly less sensitive to noise embedded in highly textured regions than in smooth, flat areas. Our framework emulates this human visual mechanism through causal disentanglement representation learning. Specifically, instead of extracting a monolithic feature where image content and distortions are mixed, our Structural Causal Model (SCM) first isolates them into independent variables. Subsequently, by
9
introducing the directed causal link C → D, we provide a structured approximation of the masking effect. This allows the network to learn how semantic and structural information modulates the visibility of distortion, ensuring that the final representation Ḋ reflects the distortion as it is actually perceived by the human eye. Connections to existing models: Existing FR-IQA models typically employ dual-branch architectures to compute feature differences or similarities between the reference and distorted images. These methods are highly effective at quantifying signal fidelity by directly comparing the paired inputs. In this work, we propose a novel and complementary perspective by formulating FR-IQA as a causal disentanglement problem. Rather than focusing primarily on pairwise feature comparisons, we explicitly model the visual masking effect. This paradigm allows us to encode how human perception of distortion is modulated by the image content.
D extracted by the encoder, the causal layer performs feature modulation in a residual form: Ḋ = D + λ (m̂ ⊙ D),
(23)
where λ is a scaling factor controlling the modulation strength. This residual modulation is important because image content should not overwrite the degradation representation itself. Instead, it should only regulate the extent to which each degradation component is expressed under a specific content condition. In this way, the causal layer treats content as an intervention variable that selectively suppresses or enhances distortion responses according to masking strength, rather than directly entangling content and degradation information. Therefore, the resulting feature Ḋ can more accurately approximate the causal effect of image content on distortion visibility, thereby reducing content-induced confounding and achieving better alignment with human subjective perception.
V. Experiments A. Model structure Figure 4 shows the overall process, including feature extraction, causal modulation, and degraded image reconstruction. Encoder: In practice, we do not perform disentanglement by explicitly splitting the features as in the theoretical analysis. Specifically, the encoder E uses a standard vision transformer vit − b − 16 to extract image degradation features. We directly feed the encoderextracted features D and the reference image C into the subsequent decoder and causal layer. This direct fusion implicitly approximates the intervention do(Z1:n−1 = C) and prevents mutual interference at the feature levels. Causal layer: As shown in Figure 10, the causal layer is designed to explicitly model the content-dependent masking effect, namely, the fact that the perceptual visibility of the same distortion varies with image content. Specifically, the feature extractor h employs a ResNet18 [42] backbone to encode the reference image C into a content representation, zc = h(C). (20) Based on zc , a nonlinear mapping f , implemented with three fully connected layers, predicts a content-aware modulation mask, m = tanh(f (zc )).
(21)
Since different image contents may produce different masking effects and thus alter the perceptual visibility of degradation components in different ways, we introduce a channel gating mechanism g and define the effective masking signal as m̂ = σ(g) ⊙ m,
(22)
where σ(·) denotes the sigmoid function and ⊙ denotes element-wise multiplication. Given the degradation feature
Fig. 10.
The structure of the decoder and causal layer.
Decoder: The reconstruction quality of distorted images is a key factor that affects the representation performance. To construct effective supervisory signals, we design a reconstruction network based on a U-Net architecture with FiLM modulation. I˙d = Decoder(Ḋ, Ir ),
(24)
where Decoder denotes the reconstruction network, Ir the reference image, and I˙d denotes the reconstructed distorted image. The overall network follows a U-Net architecture. The reference image is first mapped into feature representations by a shallow convolution and then fed into a encoder to extract multi-scale representations. The output of the encoder at layer l is denoted by El , which is used for deep semantic modeling and is also passed to the corresponding decoder layer through skip connections to preserve spatial structure and fine details. To enable the network to adaptively modulate middlefeatures according to the degradation characteristics, we
10
introduce FiLM modulation into the residual blocks of the encoder, the bottleneck, and the decoder. For an middle feature Fl ∈ RCl ×Hl ×Wl , an MLP maps the degradation feature zdeg to a channel-wise scaling parameter γl and a shifting parameter βl : γ, β = MLP(Ḋ).
(25)
We then apply a channel-wise affine transformation to the feature: FiLM(Fl , Ḋ) = (1 + γ) ⊙ Fl + β,
(26)
where ⊙ denotes channel-wise multiplication. The term (1 + γl ) ensures the stability around the identity mapping in the early stage. The architecture of the decoder is illustrated in the Figure 10. Figure 4 shows the overall process. Loss function: The final reconstruction quality of İd reflects the representation capacity of the extracted latent features. To comprehensively supervise this process, our objective function integrates three complementary components: (i) Pixel-wise Loss: We employ Mean Squared Error (MSE) between the reconstructed output İd and the target distorted image Id to enforce basic pixel-level semantic and structural alignment: n m )2 1 ∑∑(. LM SE = (27) I d (i, j) − Id (i, j) . n × m j=1 j=1 (ii) Perceptual Loss: To mitigate the over-smoothing issue inherent to MSE [43], we introduce a VGG19based perceptual loss [16], [44] to explicitly recover highfrequency structural details: ( ) LV GG = MSE VGG(I˙d ), VGG(Id ) . (28) (iii) Adversarial Loss: Finally, a PatchGAN discriminator is utilized to synthesize realistic local textures and fine-grained visual details [45]: ( ) LGAN = log (GAN(Id )) + log 1 − GAN(I˙d ) . (29)
B. Experimental details Pre-train and fine-tune: The proposed method is pretrained on a synthetic degraded dataset based on the Waterloo database. To make the proposed method adaptive to the synthetic degradations of the above IQA synthetic databases, we construct the pre-training dataset with each type of degradation at 5 degradation levels. The input images are resized to 448 × 448 pixels, randomly folded with a probability of 0.5, and normalized. AdamW [57] is used as the optimizer, while the learning rate is set to 1×10−4 initially and decreased using the cosine annealing learning rate scheduler. The epochs for pre-training and fine-tuning are 200 and 20 respectively. The fine-tuning model remains identical to those used in pre-training, with the only difference being the dataset.
Construction of Pre-training Dataset: To ensure consistency with the degradation settings of different IQA datasets, we construct the pre-training dataset according to the degradation types defined in each target dataset rather than using a unified degradation configuration. For example, for the LIVE dataset, we generate the corresponding five degradation types defined in LIVE. For other non-standard natural image datasets, we construct the pre-training data based on the degradation types described in their original dataset protocols. Predict: The UMAP is implemented based on umap − learn package. The key parameters nneighbors and min_dist are set to the default values of 15 and 0.1, respectively. PLCC and SRCC are used to evaluate model performance. The parameter sensitivity analysis and stability is provided in the supplementary materials. Datasets: We tested the proposed method on 10 IQA datasets, including five datasets that contained a wide range of distortion types (TID2013 [10], LIVE [19], CSIQ [20], KADID-10k [11], and PIPAL [12]) and five specificdomain IQA datasets related to specific content quality evaluations (Infrared image [58], Neutron image [59], Screen content image [60], Medical image [47] and Remote sensing image[]) The PLCC and SRCC as evaluation indicators. C. Quality prediction performance In this study, we evaluate a variety of representative and state-of-the-art image quality assessment techniques in five IQA datasets (LIVE IQA [35], CSIQ [46], TID2013 [34], KADID [47], and PIPAL [48]), including feature-based and deep learning approaches: PSNR, SSIM [1], MS-SSIM [49], VIF [50], VSI [51], GMSD [52], NLPD [53], PieAPP [54], LPIPS [21], DISTS [6], PDL [55], ADISTS [56], DeepWSD [18], DSD [12], TOPIQ-FR [10], DeepCausal [13], DeepJSD [14], DeepDC [15], and DBIQA [11]. The methods we compare can be categorized into three main groups: traditional image quality assessment methods, end-toend deep learning models, and training-free methods. Specifically, for the end-to-end based approaches, we utilized the weight files provided by the authors of these methods to assess their performance through cross-dataset predictions. For the training-free methods, we conducted tests directly on each dataset. For the proposed method, we evaluate performance using two prediction settings, including supervised fine-tuning and a fully unsupervised setting. Specifically, in the supervised fine-tuning setting, the model is first pretrained on a synthetic dataset, then fine-tuned for feature extraction on a standard IQA dataset, and finally used to train an MLP-based regression model with labeled IQA data for prediction. For fair comparison, we did not fine-tune on each dataset but finetuned on the PIPAL dataset for cross-dataset testing. In the PIPIL dataset, 40% of the data is used for fine-tuning and 60% for testing. In the unsupervised setting, we only pretrain the model on the synthetic dataset and then use UMPA to predict quality scores. As shown in Table I, our
11
TABLE I Results of the performance comparison conducted on benchmark datasets.
Feature Domain
LIVE [35]
Method
Traditional
Feature comparison
Causal Disentanglement
CSIQ [46]
TID2013 [34]
KADID [47]
PIPAL [48]
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PSNR SSIM [1] MS-SSIM [49] VIF [50] VSI [51] GMSD [52] NLPD [53]
0.781 0.847 0.886 0.949 0.877 0.909 0.882
0.801 0.851 0.903 0.953 0.899 0.910 0.889
0.792 0.810 0.875 0.899 0.912 0.938 0.913
0.807 0.833 0.879 0.899 0.929 0.939 0.926
0.664 0.665 0.842 0.771 0.898 0.858 0.832
0.687 0.627 0.786 0.677 0.895 0.804 0.799
0.670 0.610 0.824 0.685 0.877 0.847 0.809
0.676 0.619 0.826 0.679 0.878 0.847 0.812
0.398 0.489 0.571 0.572 0.548 0.614 0.489
0.392 0.486 0.545 0.545 0.526 0.569 0.464
PieAPP(CVPR18) [54] LPIPS(CVPR18) [21] DISTS(TPAMI20) [6] PDL [55] ADISTS [56] DeepWSD(MM22) [18] DSD [12] TOPIQ-FR(TIP24) [10] DeepCausal(CVPR25) [13] DeepJSD(TIP25) [14] DeepDC(TIP25) [15] DBIQA(TPAMI25) [11]
0.866 0.934 0.924 0.946 0.954 0.904 0.944 0.882 0.929 0.972 0.904 0.969
0.865 0.932 0.925 0.943 0.955 0.925 0.922 0.887 0.932 0.965 0.954 0.962
0.864 0.894 0.919 0.919 0.957 0.941 0.925 0.894 0.949 0.963 0.941 0.961
0.883 0.876 0.920 0.926 0.950 0.950 0.925 0.894 0.952 0.967 0.957 0.956
0.809 0.732 0.854 0.848 0.858 0.894 0.759 0.854 0.909 0.900 0.849 0.909
0.844 0.670 0.830 0.843 0.834 0.874 0.734 0.820 0.884 0.879 0.844 0.895
0.857 0.700 0.886* 0.825 0.889 0.887 0.847 0.896 0.898 0.893 0.896 0.900
0.865 0.720 0.886* 0.829 0.889 0.888 0.846 0.895 0.899 0.894 0.905 0.901
0.702 0.611 0.645 0.564 0.645 0.503 0.621 0.837* 0.675 0.612 0.669 0.745*
0.701 0.573 0.627 0.554 0.631 0.500 0.624 0.809* 0.657 0.598 0.670 0.769*
Ours Ours w/o f
0.968 0.949
0.966 0.958
0.966 0.946
0.954 0.931
0.916 0.909
0.899 0.892
0.902 0.899
0.897 0.891
0.727f 0.684
0.713f 0.666
Notably, TOPIQ-FR, DISTS, and DBIQA were trained on the KADID and PIPAL, respectively, which are marked with an asterisk (*) in the table for these methods. PieAPP and LPIPS were individually trained on their proposed datasets, respectively. (f ) represents fine-tuning in this dataset.
TABLE II Results of the performance comparison conducted on non-standard natural IQA datasets. Feature Domain
Method
Infrared [58]
Neutron [59]
Screen [60]
Medical [61]
Tone-Map [48]
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
Traditional
PSNR SSIM [1] MS-SSIM [49] VIF [50] VSI [51] GMSD [52]
0.814 0.802 0.832 0.907 0.838 0.889
0.815 0.803 0.827 0.901 0.841 0.891
0.861 0.870 0.859 0.833 0.825 0.938
0.873 0.861 0.854 0.842 0.812 0.939
0.778 0.726 0.757 0.819 0.769 0.762
0.763 0.714 0.741 0.796 0.762 0.813
0.615 0.609 0.656 0.591 0.671 0.668
0.609 0.608 0.636 0.583 0.654 0.662
-
-
Training-free
LPIPS [21] PDL [55] ADISTS [56] DeepWSD [18] DSD [12] DeepCausal [13] DeepJSD [14] DeepDC [15]
0.841 0.795 0.832 0.862 0.813 0.814 0.859 0.865
0.852 0.781 0.845 0.858 0.824 0.808 0.838 0.871
0.762 0.752 0.836 0.876 0.839 0.825 0.883 0.871
0.771 0.761 0.828 0.885 0.827 0.812 0.896 0.866
0.694 0.756 0.786 0.880 0.830 0.813 0.845 0.832
0.689 0.738 0.783 0.799 0.829 0.809 0.837 0.837
0.724 0.699 0.744 0.808 0.762 0.759 0.829 0.817
0.737 0.682 0.752 0.784 0.771 0.748 0.806 0.810
-
-
Ours
0.956
0.948
0.947
0.942
0.926
0.915
0.871
0.866
0.725
0.713
Causal Disentanglement
12
method achieves strong overall performance under both prediction settings and remains highly consistent with the ground-truth scores across all datasets. D. Cross-domain experiment on domain-specific datasets In some non-standard natural image scenarios (e.g., infrared, neutron, screen content or medical radiographic imaging), the image structure and texture differ significantly from standard natural images. Figure 11 shows images from different specific domains. Consequently, the existing sota training-free methods that rely on models pre-trained on standard natural datasets (Image-Net) are often not suitable for these scenarios. Meanwhile, obtaining authentically degraded images and annotating their quality scores in these specialized domains is notoriously difficult. Therefore, the commonly used data typically consists of subjectively annotated synthetic distorted images [59], [62]. However, constructing such IQA datasets requires a large number of expert evaluators, and the annotation process is extremely resource-intensive. The proposed method eliminates the dependency on labeled data and can be effectively pre-trained specifically for these scenarios.
construct a pre-training dataset. The proposed method is then pre-trained on this pre-training dataset, and its quality scores are predicted using UMAP. For the training-free methods, we conduct tests directly on test set using the weights provided in the corresponding papers. To ensure a rigorous and fair evaluation, the pristine images used to generate the pre-training data are strictly separated from those comprising the test set. As shown in Table II, the proposed method achieves state-of-the-art performance on all five non-standard natural IQA datasets. Some state-of-the-art training-free FRIQA methods show only suboptimal performance on these non-standard natural IQA datasets, compared with their performance on natural IQA benchmarks. For example, DeepJSD achieves state-of-the-art performance on the CISQ dataset, but its SROCC and PLCC on the neutron radiography IQA dataset are only 0.896 and 0.833, respectively. This may be because these methods typically rely on models pretrained on ImageNet, and the severe domain shift between non-standard natural images and standard natural images leads to performance degradation. Notably, some traditional FR-IQA methods also achieve strong performance on these non-standard natural image datasets. This phenomenon suggests that distortion measures based on structural, gradient, or statistical characteristics in traditional methods are often more robust in cross-domain scenarios. The strong performance of the proposed method stems from its ability to support fast training and prediction without using any labeled IQA data. In summary, the proposed method shows strong potential for practical deployment in non-standard natural image scenarios. By avoiding reliance on labeled IQA data while maintaining strong performance, it provides an effective solution for image quality assessment in specialized domains where subjective annotations are costly and difficult to obtain. E. Few-shot experiments
Fig. 11. Image examples from diverse domains. (a) to (f) are respectively natural image, infrared image, neutron radiography image, screen content image, ultrasonic image, and tone-mapped image..
To validate the adaptability and effectiveness of our approach, we evaluate a variety of representative and stateof-the-art IQA techniques across five non-standard benchmark datasets: Infrared [35], Neutron [46], Screen [34], Medical [47], and Remote sensing [48]. The compared methods include traditional metrics (PSNR, SSIM [1], MSSSIM [49], VIF [50], VSI [51], GMSD [52], NLPD [53]) and the existing SOTA training-free deep feature approaches (LPIPS [21], PDL [55], ADISTS [56], DeepWSD [18], DSD [12], DeepCausal [13], DeepJSD [14], DeepDC [15]). For the proposed method, we randomly select a subset of pristine reference images provided by the respective dataset. We then apply the dataset’s native distortion types and degradation models to these pristine images to
Unsupervised dimensionality reduction methods such as UMAP ultimately predict relative quality scores that align with human visual perception, rather than absolute scores that precisely fit MOS. Therefore, in this section, we evaluate the performance of the proposed method in predicting absolute MOS when only a small amount of IQA data is available. As shown in Table III, fewshot experiments are presented in detail on the LIVE, TID2013, KADID, and Neutron datasets. We conducted few-shot experiments on these datasets at 0%, 2%, 5%, 10%and 15% data ratios. The remaining data are all used for testing. w/frozen and w/o frozen denote the frozen/updated parameters of the degradation feature extractor. In the few-shot experiments, we use a fully connected (FC) layer with labeled IQA images for finetuning. The FC layer consists only of a normalization layer and a linear layer to map quality scores. After freezing the parameters (w/o frozen), both PLCC and SRCC stably
13
TABLE III SROCC and PLCC of different data scales across four datasets. LIVE Scale
w/ frozen
TID2013 w/o frozen
w/ frozen
KADID
w/o frozen
w/ frozen
Neutron
w/o frozen
w/ frozen
w/o frozen
PLCC SROCC PLCC SROCC PLCC SROCC PLCC SROCC PLCC SROCC PLCC SROCC PLCC SROCC PLCC SROCC 0% 2% 5% 10% 15%
0.949 0.932 0.938 0.940 0.943
0.958 0.921 0.936 0.939 0.941
0.949 0.805 0.803 0.895 0.912
0.958 0.809 0.786 0.881 0.908
0.909 0.895 0.901 0.907 0.911
0.892 0.876 0.884 0.890 0.895
0.909 0.768 0.795 0.853 0.878
0.892 0.752 0.787 0.839 0.864
0.899 0.884 0.889 0.894 0.898
0.891 0.872 0.881 0.888 0.893
TABLE IV SRCC and PLCC of different loss. Loss
LIVE
0.899 0.745 0.771 0.842 0.869
0.891 0.738 0.782 0.835 0.861
0.947 0.928 0.934 0.939 0.942
0.942 0.918 0.927 0.935 0.939
0.947 0.801 0.806 0.889 0.908
0.942 0.796 0.812 0.878 0.901
TABLE V SRCC and PLCC of different training strategies.
TID2013
Neutron
Method
LIVE
TID2013
Neutron
LM SE
LV GG
LGAN
PLCC
SRCC
PLCC
SRCC
PLCC
SRCCPT
FT
CL
PLCC
SRCC
PLCC
SRCC
PLCC
SRCC
✓ ✓ ✓ 7 ✓
7 ✓ 7 ✓ ✓
7 7 ✓ ✓ ✓
0.764 0.919 0.806 0.692 0.949
0.751 0.906 0.797 0.674 0.958
0.741 0.901 0.775 0.653 0.909
0.732 0.883 0.764 0.661 0.892
0.775 0.938 0.821 0.512 0.947
0.746 ✓ 0.927 ✓ 0.833 ✓ 0.490 7 0.942 ✓
7 ✓ 7 ✓ ✓
7 7 ✓ ✓ ✓
0.912 0.931 0.949 0.183 0.968
0.920 0.929 0.958 0.162 0.966
0.872 0.901 0.909 0.201 0.916
0.871 0.891 0.892 0.198 0.899
0.921 0.944 0.947 0.748 0.958
0.915 0.946 0.942 0.698 0.949
TABLE VI SRCC and PLCC of different projection methods.
Method KernelPCA TruncatedSVD PCA Tsne+PCA UMAP
LIVE
TID2013
TABLE VII Ablation study on causal structures.
Neutron
PLCC SRCC PLCC
SRCC PLCC
SRCC
0.937 0.930 0.942 0.897 0.949
0.859 0.861 0.906 0.862 0.892
0.921 0.919 0.937 0.862 0.942
0.926 0.934 0.955 0.889 0.958
0.871 0.869 0.916 0.891 0.909
0.934 0.926 0.951 0.884 0.947
maintain high values, demonstrating that the extracted degradation features exhibit strong generalization ability and excel in zero-shot and few-shot scenarios. However, when only a small amount of labeled IQA data is used to fully fine-tune the parameters of the degradation feature extractor, these data fail to provide effective supervision and instead damage the extracted features, leading to a severe performance drop. This indicates that a small amount of labeled IQA data is insufficient to align the model more precisely with human subjective perception and tends to degrade the original feature representation. F. Ablation Studies The impact of loss: During pre-training and fine-tuning, we employ three loss functions: MSE loss, VGG loss, and GAN loss. Pixel-wise loss functions such as MSE struggle to handle the uncertainty inherent in recovering lost high-frequency details such as texture. When solely MSE loss is used, the model fails to reconstruct images degraded with high-frequency noise, which leads to the model lacking the ability to capture high-frequency degradation. Across all three datasets, introducing LV GG on top of LM SE yields the largest single-step gain, which indicates that effective degradation-aware supervision is
Configuration Ours w/o Causal Inverse Causal-VAE VGG16 ResNet18
Causal Graph C⊥D D→C C→D C→D C→D
LIVE
TID2013
Neutron
PLCC
SRCC PLCC SRCC PLCC
SRCC
0.912 0.908 0.923 0.941 0.949
0.920 0.916 0.938 0.939 0.958
0.915 0.914 0.921 0.929 0.942
0.872 0.865 0.895 0.896 0.909
0.871 0.869 0.878 0.885 0.892
0.921 0.917 0.925 0.933 0.947
the dominant factor in improving representation quality. Specifically, when moving from using only LM SE to using LM SE + LV GG , the PLCC/SRCC on LIVE increases from 0.764/0.751 to 0.919/0.906, that on TID2013 increases from 0.741/0.732 to 0.901/0.883, and that on Neutron increases from 0.775/0.746 to 0.938/0.927. Figure 12 presents the scatter plot that evaluates high-frequency and low-frequency degradations with only MSE loss. The proposed model performs well on evaluating certain lowfrequency degradation types, but it lacks the capability to assess high-frequency degradation types. The adversarial loss primarily improves the fidelity of fine details and textures in the reconstructed images, thereby enhancing the representation capability of the model. The full loss combination achieves the best performance on all datasets, reaching 0.949/0.958 on LIVE, 0.909/0.892 on TID2013, and 0.947/0.942 on Neutron. Compared with LM SE +LV GG , the improvement is smaller, which suggests that LGAN mainly serves as a further refinement term. The Impact of training strategies: We investigate the contributions of three training strategies: pre-training (PT), fine-tuning (FT), and the causal layer (CL). As shown in Table V, PT equips the model with a basic ca-
14
likely stems from the difference in modeling focus: CausalVAE is designed for generic disentanglement of multiple latent factors, whereas our method explicitly models the content–degradation relationship that is central to IQA. TABLE VIII UMAP Parameter Sensitivity and Stability Experiment Results (TID2013 Dataset, 10 Repeats)
(a) Low frequency
(b) High frequency
Fig. 12. Scatter plot of partial low-frequency degradation assessment versus high-frequency degradation assessment.
pability to disentangle degradation features.. Even the use of only the pre-training scheme yields strong performance on the LIVE, TID2013, and Neutron datasets. CL captures the causal effect of image content on degradation-related features. With this layer, the learned representation is better aligned with human subjective perception in an unsupervised setting, leading to consistent performance gains across all three datasets. After labeled data are introduced during the final stage of FT, the model is more precisely aligned with human subjective perception and achieves the best performance on all three datasets. The impact of projection methods: We tested the impact of different dimensionality reduction methods (TSNE [63], Kernel PCA [64], TruncatedSVD [65], PCA [66]) and UMAP [67]. As shown in Table VI, different projection methods significantly influence the performance of the proposed method, with UMAP achieving the best performance. In the ablation experiment, the performance of TSNE fluctuates with random seeds, while other projection methods usually demonstrate stable results. All methods use their default parameters. Performance differences stem from the differing principles of dimensionality reduction methods. The nonlinear dimensionality reduction method based on manifold learning (UMAP) balances local and global structures. T-SNE may overly focus on local structures while neglecting global features, leading to inferior IQA performance. The impact of causal layer: Table VII reports an ablation study on causal-layer design choices. Specifically, the causal structure between content and degradation is not interchangeable. Compared with the variant without causal modeling (C ⊥ D), our full model with C → D achieves clear and consistent gains on all datasets. In contrast, reversing the direction to D → C results in inferior performance, indicating that the proposed C → D formulation better matches the underlying dependency in distorted images. The choice of backbone substantially affects performance. When the same causal graph (C → D) is used, ResNet18 consistently surpasses VGG16 across LIVE, TID2013, and Neutron, yielding the strongest PLCC and SRCC in every case. This demonstrates that the proposed causal layer benefits from a more expressive feature extractor. The ResNet18/VGG16-based architecture also outperform Causal-VAE [26]. This improvement
Type
Value
PLCC (Mean±Std)
SROCC (Mean±Std)
Varying nneighbors (fixed min_dist = 0.1) nneighbors nneighbors nneighbors
5 15 (default) 30
0.887 ± 0.003 0.909 ± 0.004 0.901 ± 0.003
0.872 ± 0.002 0.892 ± 0.002 0.886 ± 0.003
Varying min_dist (fixed nneighbors = 15) min_dist min_dist min_dist
0.01 0.1 (default) 0.5
0.893 ± 0.002 0.909 ± 0.003 0.905 ± 0.003
0.878 ± 0.003 0.892 ± 0.004 0.892 ± 0.002
Parameter Sensitivity and Stability Experiments: To verify the parameter sensitivity and stability of UMAP in quality score prediction, we conduct experiments on the TID2013 dataset by varying two key parameters— nneighbors and min_dist—with 10 repeated runs for each parameter combination. As shown in Table VIII, the PLCC and SROCC metrics fluctuate within ±0.004 across all tested parameter settings, demonstrating excellent stability and robustness of our method to parameter variations. The default parameters (nneighbors = 15, min_dist = 0.1) yield optimal performance, and even with reasonable parameter adjustments, the performance degradation is negligible. This confirms the reliability of UMAP for projecting high-dimensional degradation features into a 1-D space for quality scoring. G. Visualization The reconstruction quality of distorted images is a key factor that affects the representation performance. As shown in Figure 13, the top row presents the distorted images, the left column presents the reference images, and the bottom row presents the reconstructed distorted images. The results show that the proposed method reconstructs the distorted images with high fidelity, which provides effective supervision signals for the model. VI. Limitations and failure cases Limitations: Although our method achieves strong performance with domain-specific pre-training, this dependence also limits its applicability. Unlike fully training-free FR-IQA methods, it still requires clean reference data and explicit degradation simulation to build the pre-training set. As a result, it does not offer the same plug and play convenience as fully training-free approaches. Failure cases: The proposed method achieves strong performance on common global distortions, such as blur, noise, and compression, but performs relatively worse on blockwise distortions. This limitation stems from two factors. Blockwise distortions disrupt the original image
15
Fig. 13. Example results of the proposed method. For instance, given a distorted image and a reference image, the proposed method attempts to transfer the degradations to the reference image.
structure, which makes accurate reconstruction difficult (see Figure 13). In this case, the reconstruction losses, including LM SE, LV GG, and LGAN , cannot provide effective supervision for learning reliable degradation representations. Blockwise distortions also directly damage the original image content, which conflicts with the visual masking effect. As a result, the proposed method cannot reliably capture the interaction between content and degradation, which limits its prediction performance. VII. Conclusions Inspired by the concept of decoupled representation learning, we propose an FR-IQA method based on causal inference and decoupled representation learning. Specifically, the method constructs a model that decouples content-influenced degradation features. Subsequently, the quality score is predicted by learning the approximate manifold of the degradation features. Experimental results verify the effectiveness of our method. Moreover, the method demonstrates superior cross-domain generalization compared with existing training-free FR-IQA models, owing to its capability for scenario-specific training without labeled IQA data. VIII. ACKNOWLEDGEMENT This work is supported by the National Natural Science Foundation of China (Nos. 62276218, 12192214, 62373096), the Fundamental Research Funds for the Central Universities, China (No. 2682024ZTPY055), Science and Technology Development Plan Project of Jilin Province, China. Grant No. 20260206058ZP, and National Natural Science Foundation of China (Grant No. 11905028). References [1] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. 1, 10, 11, 12 [2] L. Zhang, L. Zhang, X. Mou, and D. Zhang, “Fsim: A feature similarity index for image quality assessment,” IEEE Transactions on Image Processing, vol. 20, no. 8, pp. 2378–2386, 2011. 1 [3] A. Saha and Q. J. Wu, “Full-reference image quality assessment by combining global and local distortion measures,” Signal Processing, vol. 128, pp. 186–197, 2016. 1
[4] A. Rehman and Z. Wang, “Reduced-reference image quality assessment by structural similarity estimation,” IEEE Transactions on Image Processing, vol. 21, no. 8, pp. 3378–3389, 2012. 1 [5] A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,” IEEE Transactions on Image Processing, vol. 21, no. 12, pp. 4695–4708, 2012. 1 [6] K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 5, pp. 2567–2581, 2020. 1, 3, 10, 11 [7] ——, “Comparison of full-reference image quality models for optimization of image processing systems,” International Journal of Computer Vision, vol. 129, no. 4, pp. 1258–1281, 2021. 1 [8] X. Liao, X. Wei, M. Zhou, Z. Li, and S. Kwong, “Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces,” IEEE Transactions on Image Processing, vol. 33, pp. 4044–4059, 2024. 1 [9] W. Zhang, D. Li, X. Min, G. Zhai, G. Guo, X. Yang, and K. Ma, “Perceptual attacks of no-reference image quality models with human-in-the-loop,” Advances in Neural Information Processing Systems, vol. 35, pp. 2916–2929, 2022. 1 [10] C. Chen, J. Mo, J. Hou, H. Wu, L. Liao, W. Sun, Q. Yan, and W. Lin, “Topiq: A top-down approach from semantics to distortions for image quality assessment,” IEEE Transactions on Image Processing, vol. 33, pp. 2404–2418, 2024. 1, 3, 10, 11 [11] X. Liao, X. Wei, M. Zhou, H.-S. Wong, and S. Kwong, “Image quality assessment: Exploring joint degradation effect of deep network features via kernel representation similarity analysis,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1, 2, 3, 10, 11 [12] I. Kligvasser, T. Shaham, Y. Bahat, and T. Michaeli, “Deep self-dissimilarities as powerful visual fingerprints,” Advances in Neural Information Processing Systems, vol. 34, pp. 3939–3951, 2021. 2, 10, 11, 12 [13] W. Shen, M. Zhou, Y. Chen, X. Wei, Y. Feng, H. Pu, and W. Jia, “Image quality assessment: Investigating causal perceptual effects with abductive counterfactual inference,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 990–17 999. 2, 3, 10, 11, 12 [14] X. Liao, X. Wei, M. Zhou, Z. Li, and S. Kwong, “Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces,” IEEE Transactions on Image Processing, vol. 33, pp. 4044–4059, 2024. 2, 10, 11, 12 [15] H. Zhu, B. Chen, L. Zhu, S. Wang, and W. Lin, “Deepdc: Deep distance correlation as a perceptual image quality evaluator,” IEEE Transactions on Image Processing, vol. 34, pp. 7859–7873, 2025. 2, 3, 10, 11, 12 [16] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. 2, 10 [17] M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” in International conference on machine learning. PMLR, 2019, pp. 6105–6114. 2 [18] X. Liao, B. Chen, H. Zhu, S. Wang, M. Zhou, and S. Kwong, “Deepwsd: Projecting degradations in perceptual space to wasserstein distance in deep feature space,” in Proceedings of
16
the 30th ACM International Conference on Multimedia, 2022, pp. 970–978. 2, 3, 10, 11, 12 [19] F. Gao, Y. Wang, P. Li, M. Tan, J. Yu, and Y. Zhu, “Deepsim: Deep similarity for image quality assessment,” Neurocomputing, vol. 257, pp. 104–114, 2017. 3 [20] J. Kim and S. Lee, “Deep learning of human visual sensitivity in image quality assessment framework,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1676–1684. 3 [21] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595. 3, 10, 11, 12 [22] K. Ding, Y. Liu, X. Zou, S. Wang, and K. Ma, “Locally adaptive structure and texture similarity for image quality assessment,” in Proceedings of the 29th ACM International Conference on multimedia, 2021, pp. 2483–2491. 3 [23] H. D. Sikka, A Deeper Look at the Unsupervised Learning of Disentangled Representations in β-VAE from the perspective of Core Object Recognition. Harvard University, 2020. 3, 6 [24] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013. 3 [25] I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in International conference on learning representations, 2017. 3 [26] M. Yang, F. Liu, Z. Chen, X. Shen, J. Hao, and J. Wang, “Causalvae: Disentangled representation learning via neural structural causal models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9593–9602. 3, 4, 6, 14 [27] H. Nam, “Scadi: Self-supervised causal disentanglement in latent variable models,” arXiv preprint arXiv:2311.06567, 2023. 3, 4, 5, 6 [28] A. J. Ahumada Jr and H. A. Peterson, “Luminance-model-based dct quantization for color image compression,” in Human vision, visual processing, and digital display III, vol. 1666. SPIE, 1992, pp. 365–374. 3, 8 [29] D. G. Pelli and P. Bex, “Measuring contrast sensitivity,” Vision research, vol. 90, pp. 10–14, 2013. 3, 8 [30] G. E. Legge and J. M. Foley, “Contrast masking in human vision,” Journal of the optical Society of America, vol. 70, no. 12, pp. 1458–1471, 1980. 3, 8 [31] Z. Chen and C. Guillemot, “Perceptually-friendly h. 264/avc video coding based on foveated just-noticeable-distortion model,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 6, pp. 806–819, 2010. 3, 8 [32] X. Wang, J. Xiong, and W. Lin, “Visual interaction perceptual network for blind image quality assessment,” IEEE Transactions on Multimedia, vol. 25, pp. 8958–8971, 2023. 3 [33] L. Zheng, Y. Luo, Z. Zhou, J. Ling, and G. Yue, “Cdinet: Content distortion interaction network for blind image quality assessment,” IEEE Transactions on Multimedia, vol. 26, pp. 7089–7100, 2024. 3 [34] N. Ponomarenko, L. Jin, O. Ieremeiev, V. Lukin, K. Egiazarian, J. Astola, B. Vozel, K. Chehdi, M. Carli, F. Battisti et al., “Image database tid2013: Peculiarities, results and perspectives,” Signal Processing: Image Communication, vol. 30, pp. 57–77, 2015. 5, 10, 11, 12 [35] H. R. Sheikh, M. F. Sabir, and A. C. Bovik, “A statistical evaluation of recent full reference image quality assessment algorithms,” IEEE Transactions on Image Processing, vol. 15, no. 11, pp. 3440–3451, 2006. 5, 10, 11, 12 [36] X. Shen, F. Liu, H. Dong, Q. Lian, Z. Chen, and T. Zhang, “Weakly supervised disentangled generative causal representation learning,” Journal of Machine Learning Research, vol. 23, no. 241, pp. 1–55, 2022. 5, 6 [37] F. Locatello, S. Bauer, M. Lucic, G. Raetsch, S. Gelly, B. Schölkopf, and O. Bachem, “Challenging common assumptions in the unsupervised learning of disentangled representations,” in international conference on machine learning. PMLR, 2019, pp. 4114–4124. 5 [38] J. Pearl, Causality. Cambridge university press, 2009. 6 [39] M. J. Kusner, J. Loftus, C. Russell, and R. Silva, “Counterfactual fairness,” Advances in neural information processing systems, vol. 30, 2017. 6
[40] Y. Liu, J. He, J. Gu, X. Kong, and C. Dong, “Degae: A new pretraining paradigm for low-level vision,” Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pp. 23 292–23 303, 2023. 6 [41] N.-H. Shin, S.-H. Lee, and C.-S. Kim, “Blind image quality assessment based on geometric order learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 12 799–12 808. 8 [42] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778. 9 [43] J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in European conference on computer vision. Springer, 2016, pp. 694–711. 10 [44] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang et al., “Photo-realistic single image super-resolution using a generative adversarial network,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4681–4690. 10 [45] M. Tahmid, M. S. Alam, N. Rao, and K. M. A. Ashrafi, “Imageto-image translation with conditional adversarial networks,” in 2023 IEEE 9th International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECONECE), 2023, pp. 1–5. 10 [46] E. C. Larson and D. M. Chandler, “Most apparent distortion: full-reference image quality assessment and the role of strategy,” Journal of electronic imaging, vol. 19, no. 1, pp. 011 006–011 006, 2010. 10, 11, 12 [47] H. Lin, V. Hosu, and D. Saupe, “Kadid-10k: A large-scale artificially distorted iqa database,” in 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX). IEEE, 2019, pp. 1–3. 10, 11, 12 [48] G. Jinjin, C. Haoming, C. Haoyu, Y. Xiaoxing, J. S. Ren, and D. Chao, “Pipal: a large-scale image quality assessment dataset for perceptual image restoration,” in European conference on computer vision. Springer, 2020, pp. 633–651. 10, 11, 12 [49] Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” in The thrityseventh asilomar conference on signals, systems & computers, 2003, vol. 2. Ieee, 2003, pp. 1398–1402. 10, 11, 12 [50] H. R. Sheikh and A. C. Bovik, “Image information and visual quality,” IEEE Transactions on image processing, vol. 15, no. 2, pp. 430–444, 2006. 10, 11, 12 [51] L. Zhang, Y. Shen, and H. Li, “Vsi: A visual saliency-induced index for perceptual image quality assessment,” IEEE Transactions on Image processing, vol. 23, no. 10, pp. 4270–4281, 2014. 10, 11, 12 [52] W. Xue, L. Zhang, X. Mou, and A. C. Bovik, “Gradient magnitude similarity deviation: A highly efficient perceptual image quality index,” IEEE transactions on image processing, vol. 23, no. 2, pp. 684–695, 2013. 10, 11, 12 [53] V. Laparra, J. Ballé, A. Berardino, and E. P. Simoncelli, “Perceptual image quality assessment using a normalized laplacian pyramid,” Electronic Imaging, vol. 28, pp. 1–6, 2016. 10, 11, 12 [54] E. Prashnani, H. Cai, Y. Mostofi, and P. Sen, “Pieapp: Perceptual image-error assessment through pairwise preference,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1808–1817. 10, 11 [55] M. Delbracio, H. Talebei, and P. Milanfar, “Projected distribution loss for image enhancement,” in 2021 IEEE International Conference on Computational Photography (ICCP). IEEE, 2021, pp. 1–12. 10, 11, 12 [56] K. Ding, Y. Liu, X. Zou, S. Wang, and K. Ma, “Locally adaptive structure and texture similarity for image quality assessment,” in Proceedings of the 29th ACM International Conference on multimedia, 2021, pp. 2483–2491. 10, 11, 12 [57] I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017. 10 [58] O. Zelmati, B. Bondžulić, B. Pavlović, I. Tot, and S. Merrouche, “Study of subjective and objective quality assessment of infrared compressed images,” Journal of Electrical Engineering, vol. 73, no. 2, pp. 73–87, 2022. 10, 11 [59] Z. Zhang, C.-B. Meng, X.-L. Jiang, C.-Y. Zhao, S. Qiao, and T. Zhang, “Comprehensive quality assessment method for
17
neutron radiographic images based on cnn and visual salience,” Nuclear Science and Techniques, vol. 36, no. 7, p. 118, 2025. 10, 11, 12 [60] Z. Ni, L. Ma, H. Zeng, Y. Fu, L. Xing, and K.-K. Ma, “Scid: A database for screen content images quality assessment,” in 2017 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS). IEEE, 2017, pp. 774–779. 10, 11 [61] S. Zhang, Y. Wang, J. Jiang, J. Dong, W. Yi, and W. Hou, “Cnn-based medical ultrasound image quality assessment,” Complexity, vol. 2021, no. 1, p. 9938367, 2021. 11 [62] Q. Zhao, H. Liu, W. Tian, and X. Wang, “Construction of the hyperspectral image distortion evaluation index for low altitude uavs,” Trans. Chin. Soc. Agric. Eng, vol. 20, pp. 67–76, 2022. 12 [63] L. Van der Maaten and G. Hinton, “Visualizing data using tsne.” Journal of Machine Learning Research, vol. 9, no. 11, 2008. 14 [64] B. Schölkopf, A. Smola, and K.-R. Müller, “Kernel principal component analysis,” in International Conference on Artificial Neural Networks. Springer, 1997, pp. 583–588. 14 [65] A. Falini, “A review on the selection criteria for the truncated svd in data science applications,” Journal of Computational Mathematics and Data Science, vol. 5, p. 100064, 2022. 14 [66] A. Maćkiewicz and W. Ratajczak, “Principal components analysis (pca),” Computers & Geosciences, vol. 19, no. 3, pp. 303–342, 1993. 14 [67] L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,” arXiv preprint arXiv:1802.03426, 2018. 14 Zhen Zhang is pursuing the Ph.D. degree in the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China. His research interests include Causal decoupling representation learning, causal inference, continual learning, and image quality assessment.
Jielei Chu (Senior Member, IEEE) received the Ph.D. degree in computer science from Southwest Jiaotong University, Chengdu, China, in 2020. He serves as an Editorial Board Member for Scientific Reports. He has published more than 30 papers in journals such as IEEE TPAMI, IEEE TKDE, IEEE TCYB, and IEEE TMM. His research interests include deep learning, semi-supervised learning, federated learning, and brain-inspired intelligence.
Tian Zhang (Member, IEEE) received the Ph.D. degree in electronic circuit and systems from the State Key Laboratory of Integrated Optoelectronics, College of Electronic Science and Engineering, Jilin University, Changchun, in 2016. He is an associate professor at Northeast Normal University. He serves as Editor members of Scientific Reports from 2025. He has authored more than 40 technical publications and 6 patents. His current research interests include miniaturized neutron radiography equipment and neutron radiography image processing. Weide Liu (Member, IEEE) received the bachelor’s and Ph.D. degrees from Nanyang Technological University. He is currently a Research Fellow with Boston Children’s Hospital, Harvard Medical School. Previously, he held the position of a Research Scientist with A*STAR, Singapore. His research interests include computer vision, natural language processing, machine learning, and medical image analysis.
Fengmao Lv (Member, IEEE) received the B.S. and Ph.D. degrees in computer science from the University of Electronic Science and Technology of China, Chengdu, China, in 2013 and 2018, respectively. He is currently an Associate Professor with Southwest Jiaotong University. He serves as an Associate Editor for IEEE Transactions on Circuits and Systems for Video Technology, Action Editor for Neural Networks, and Area Chair for ACM MM. His research interests include multimodal learning, open-world learning, and multimedia analysis.
Tianrui Li (Senior Member, IEEE) received the B.S., M.S., and Ph.D. degrees from Southwest Jiaotong University, Chengdu, China, in 1992, 1995, and 2002, respectively. He is currently a Professor and the Director of the Key Laboratory of Cloud Computing and Intelligent Techniques, Southwest Jiaotong University. He serves as Editor-in-Chief of HumanCentric Intelligent Systems, Editor of Information Fusion, and Associate Editor of ACM TIST. He has authored or coauthored more than 500 papers in journals and conferences such as IEEE TPAMI, IJCV, CVPR, and ICCV. His research interests include big data, data mining, cloud computing, granular computing, and rough sets. He is a Fellow of IRSS and a Senior Member of ACM and IEEE.
Jun Cheng (Senior Member, IEEE) received the B.E. degree in electronic engineering and information science from the University of Science and Technology of China, and the Ph.D. degree from Nanyang Technological University, Singapore. He is currently a Principal Research Scientist with the Institute for Infocomm Research, A*STAR, working on AI for medical imaging, robust machine vision, and perception. Dr. Cheng is an Associate Editor of IEEE Transactions on Medical Imaging and a Senior Area Editor of IEEE Transactions on Image Processing.
Yuming Fang (Fellow, IEEE) received the B.E. degree from Sichuan University, Chengdu, China, the M.S. degree from the Beijing University of Technology, Beijing, China, and the Ph.D. degree from Nanyang Technological University, Singapore. He is currently a Professor with the School of Information Management, Jiangxi University of Finance and Economics, Nanchang, China. His research interests include visual attention modeling, visual quality assessment, computer vision, and 3-D image/video processing. He is on the Editorial Board for IEEE Transactions on Multimedia and Signal Processing: Image Communication.