Perceptual Decoupling With a New Dataset and the HiFi-Net Blind Image Quality Assessment Model
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Traditional Image Quality Assessment (IQA) often treats perceptual quality as a monolithic concept, overlooking the independent contributions of its color and spatial dimensions. To address this gap, this research first introduces a large-scale subjective study to investigate the relationship between total image quality (IQT) and its constituent color (IQC) and spatial (IQS) components. The results provide strong empirical evidence that IQTcan be effectively decoupled and accurately represented by a weighted linear combination of IQCand IQS(Pearson Correlation Coefficient of 0.96), while IQCand IQSthemselves are largely independent (correlation of 0.092). Theoretically, this perceptual decoupling aligns with the functional specialization of the Human Visual System (HVS), which processes chromatic and structural attributes through segregated neural pathways. Moreover, we propose a novel Blind IQA (BIQA) model, the Hierarchical Fusion Network (HiFi-Net). This model leverages a pre-trained Vision Transformer (ViT) backbone to extract rich, multi-level features and introduces an innovative layer-independent perceptual attention mechanism (comprising intra-layer and inter-layer attention). This mechanism adaptively captures layer-specific quality-aware features and aggregates them across depths, extracting critical information from complex distortions for final score regression. Extensive experiments on nine public benchmark datasets demonstrate that HiFi-Net achieves state-of-the-art performance on both synthetic and authentic distortions, validating its effectiveness and strong generalization capability.
Fields of Study & Subjects
Image and Video Quality Assessment
Digital Media Forensic Detection
Image Retrieval and Classification Techniques
Publication details
- Journal / source
- IEEE Access
- Volume
- 14
- Pages
- 51049–51061
- Publication date
- 1 January 2026
- Language
- EN