Adaptive Spectral Reweighting via Distortion Prior for No‐Reference Image Quality Assessment
Si Chen, Menglong Chen, Yarong Sun, Yukuan Sun, Xuesong Li, Jing Liu, Xiaojie DuanABSTRACT
No‐reference image quality assessment (NR‐IQA) aims to predict perceptual image quality without access to reference images. From a signal processing perspective, image distortions can be interpreted as spectral redistribution operators that alter the energy distribution across frequency bands. However, the impact of distortions on spectral representations is not uniform. Different distortions affect frequency components in different ways, and some distortions introduce only weak or even negligible spectral changes. This suggests that frequency‐domain information is not equally informative for all distortion types, whereas most existing NR‐IQA methods still adopt shared, distortion‐agnostic strategies for frequency feature modelling. Inspired by this observation, we propose the frequency and distortion integrated quality network (FDIQN), a distortion‐prior‐guided framework for adaptive frequency feature extraction. Specifically, a distortion perception module (DPM) is first introduced to infer latent distortion priors directly from the input image. These priors are then used to guide a distortion‐conditioned frequency attention (DCFA) mechanism, which adaptively emphasizes informative DCT‐based frequency components according to the underlying degradation characteristics. In this way, the proposed model can selectively exploit frequency‐domain cues when they are truly relevant to perceptual quality prediction, instead of applying a unified spectral processing scheme to all distortions. Extensive experiments on six benchmark IQA datasets demonstrate that FDIQN achieves competitive and consistently strong performance across both synthetic and authentic distortion scenarios. Additional cross‐dataset ablation studies and distortion‐wise analyses further verify the effectiveness of the proposed distortion‐guided frequency modelling strategy, highlighting the importance of explicitly modelling distortion‐dependent spectral characteristics for NR‐IQA.