Image Quality Assessment: Unifying Spatial and Frequency Distribution Discrepancy in Deep Feature Domains via Rényi Divergence

Full-reference image quality assessment (FR-IQA) serves as a fundamental component in computer vision tasks, yet most existing approaches are constrained by the information degradation introduced through hand-designed feature representations, which prevents them from performing accurate assessments. By comparing differences in probability distributions between images, FR-IQA models can reduce reliance on specific visual features to some extent, which raises the issue of ensuring consistency between distributional differences and human perception. In response to these challenges, this study proposes a deep feature-based FR-IQA framework that combines spatial and frequency domain distribution comparisons, thereby avoiding reliance on specific visual features while aligning with human visual perception. First, we obtain the distribution representation of the images in the deep feature domain through a pretrained backbone network, thereby avoiding reliance on specific visual information and effectively measuring the image distribution differences via Rényi divergence. Second, we extend the distribution measurement to both the spatial and frequency domains to balance the consistency of visual perception with the accuracy of the difference comparison, subsequently obtaining the final quality score through a cross-domain mapping strategy. Third, the proposed framework eliminates the need for elaborate feature engineering and complex model training, ensuring interpretability from both statistical and perceptual perspectives, which demonstrates state-of-the-art performance and robust generalizability through extensive experiments.