Analysis of PSNR, SSIM, LPIPS metrics in the context of human perception of visual similarity

Abstract

This paper presents a comprehensive comparative analysis of three well-known image quality assessment (IQA) metrics: PSNR, SSIM, and LPIPS. It explores their basic principles, mathematical foundations, advantages, and limitations, particularly as they relate to human visual perception. The evolution of IQA metrics from simple pixel-by-pixel comparisons (PSNR) to structural approaches (SSIM) and, more recently, to learned perceptual metrics (LPIPS) is discussed. A critical analysis of the effectiveness of each metric in assessing various visual distortions, including noise, blur, and compression artifacts, is presented. Inherent issues in human visual perception, such as the role of semantics, texture, color, and visual artifacts, are explored as fundamental causes of discrepancies between objective metric estimates and subjective human judgments. The paper highlights the “unproven effectiveness” of deep features in LPIPS, and discusses its vulnerabilities, such as adversarial attacks and limitations in global semantic understanding. Finally, it outlines directions for future research aimed at developing more robust, interpretable, and perceptually consistent IQA metrics that can better account for the complexity of the human visual system and the evolving demands of modern image processing and generative artificial intelligence technologies.