Structure-Aware Artifact Peak Signal-to-Noise Ratio for Image Quality Assessment

Current image quality assessment (IQA) methods generally adopt a rigid fidelity criterion that penalizes all deviations of the distorted image from the reference image, whether they are beneficial enhancements or harmful artifacts. The recently proposed artifact peak signal-to-noise ratio (APSNR) overcomes this limitation by introducing a selective fidelity criterion, which allows legitimate enhancements while penalizing only spurious artifacts. However, APSNR identifies artifacts through a global distorted-reference mapping, which may misclassify locally varying enhancements as artifacts, such as non-uniform illumination variations, highlight suppression, or texture smoothing.To address this challenge, we propose the structure-aware APSNR (SAAPSNR), a novel IQA method based on the selective fidelity criterion that characterizes artifacts through local structural deviations. SAAPSNR first represents the local structure around each pixel in tensor form. It then traverses all corresponding neighborhoods between the distorted and reference images to compute the artifact response, generating an artifact response map. Finally, the artifact responses are aggregated into a score to assess the overall image quality. Extensive experiments demonstrate that SAAPSNR closely aligns with human perception of artifacts, while offering increased generality in accommodating diverse image processing pipelines and enhancement strategies, making it a reliable benchmark tool for evaluating artifact suppression in image processing.