No-Reference Image Quality Assessment via Perception-Guided Distortion Representation Refinement

Abstract

No-Reference Image Quality Assessment (NR-IQA) aims to predict perceptual image quality from distorted images without reference signals. Existing NR-IQA methods often incorporate visual attention through external weighting or feature fusion, while semantic and distortion cues are often not sufficiently organized for quality prediction, limiting perceptually guided distortion refinement. To address these limitations, this paper proposes a perception-guided distortion representation refinement framework for NR-IQA. Specifically, the Perceptual Adaptive Network (PAN) converts visual attention priors into spatially varying affine parameters for feature normalization, enabling perceptual-prior-conditioned modulation of intermediate distortion responses. The Semantic-distortion Asymmetric Cross-Gated Head (SD-ACG) further organizes the quality representation into semantic-aware and distortion-sensitive branches, using semantic-guided residual cross-gating and orthogonal branch regularization to refine distortion representations while reducing branch redundancy. Experiments on standard synthetic and authentic IQA benchmarks demonstrate consistent performance improvements across diverse distortion settings.