A new image similarity metric for a perceptual and transparent geometric and chromatic assessment

Abstract

Full-reference image quality assessment metrics such as PSNR and SSIM misalign with human perceptual judgments on texture and shape distortions, a limitation documented across multiple benchmarks. Deep perceptual metrics such as LPIPS substantially close this gap but produce opaque scores that cannot be attributed to specific image regions, a concern amplified by emerging Artificial Intelligence regulation requirements. We propose EDOKS (Earth mover’s Distance and Oklab Similarity), a white-box full-reference IQA metric that decomposes perceptual dissimilarity into two interpretable terms: a texture term, computed via Earth Mover’s Distance on Gabor filter signatures, and a chromatic term, computed as pixel-wise Euclidean distance in the Oklab perceptual color space. Evaluated on the Berkeley-Adobe Perceptual Patch Similarity dataset, which includes traditional, CNN-generated, and geometric distortions, EDOKS outperforms all tested low-level full-reference competitors on both the JND and 2AFC subsets, measured respectively by correlation with human judgments and agreement with human choices, achieving results close to deep perceptual metrics without any learned parameters. Beyond the scalar score, EDOKS produces pixel-level heatmaps that localize texture and chromatic discrepancies independently, exposing failure modes invisible to luminance-based metrics and providing users with verifiable evidence for the assigned similarity score.