AI RESEARCH BREAKDOWN

Causal Disentanglement for Full-Reference Image Quality Assessment

Existing deep network-based full-reference image quality assessment (FR-IQA) models typically rely on pairwise comparisons of deep features from reference and distorted images. Although effective on standard IQA benchmarks, existing FR-IQA methods still face two major limitations. Training-dependent methods require labeled IQA data for supervised optimization, but reliable quality annotations are difficult to obtain because they rely on subjective experiments. Training-free methods avoid supervised training, but their fixed perceptual priors limit their adaptability to non-standard or domain-specific image scenarios. To address these limitations, we propose a novel FR-IQA paradigm based on causal disentanglement representation learning. Unlike conventional feature comparison-based methods, our approach formulates degradation estimation as a causal disentanglement process guided by interventions on latent representations. Specifically, we first decouple degradation and content representations by exploiting the content invariance between reference and distorted images. Inspired by the human visual masking effect, we then design a masking module to model the causal influence of image content on degradation features, thereby extracting content-influenced degradation representations from distorted images. Finally, quality scores are predicted from these representations using either supervised regression or label-free dimensionality reduction. Extensive experiments show that our method achieves highly competitive performance on standard IQA benchmarks under fully supervised, few-label, and label-free settings. Moreover, on diverse non-standard image domains with scarce data, including infrared, neutron, screen-content, medical, and tone-mapped images, our method exhibits stronger MOS-free domain adaptation than existing training-free FR-IQA models.