PSIM: Perceptual Similarity Index Measure

Md Eimran Hossain Eimon, Hari Kalva; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026, pp. 8564-8574

Human perception integrates information across multiple spatial scales, rapidly detecting coarse, appearance-level distortions while relying on high-acuity mechanisms to identify near-threshold deviations. Existing full-reference image quality assessment (FR-IQA) models often fail to capture this balance: some emphasize global structure at the expense of fine detail, while others excel at local fidelity but overlook global perceptual changes. To address this gap, we introduce Perceptual Similarity Index Measure (PSIM), a perceptually motivated FR-IQA model that utilizes an advanced Multi-level Wasserstein Distortion (MWD) model. MWD evaluates image distortions across a hierarchy of spatial supports, enabling the model to capture both Most Apparent Distortions (MAD), which are large-scale and visually dominant changes that attract immediate attention, and Least Apparent Distortions (LAD), which are subtle, fine-grained deviations that become noticeable only under closer inspection. By computing Wasserstein distortions over progressively larger receptive fields, PSIM provides a unified representation of both coarse and fine perceptual regimes. Comprehensive evaluations across multiple public IQA datasets demonstrate that PSIM outperforms most existing state-of-the-art metrics while requiring significantly fewer FLOPs. Moreover, it generalizes strongly in cross-dataset evaluations and exhibits robustness to small spatial shifts. Further analysis demonstrates that PSIM achieves competitive performance in assessing perceptual color differences.