A Comprehensive Framework for Perceptual Quality Assessment in AI-Based Image Compression
Abstract
Perceptual quality assessment has become the central bottleneck in the design, evaluation, and deployment of artificial-intelligence-based image compression systems. Classical signal-fidelity measures such as mean squared error (MSE) and peak signal-to-noise ratio (PSNR) correlate poorly with human judgment. Learned codecs, meanwhile, increasingly optimize directly for structural, perceptual, and generative objectives. How those objectives behave under the rate–distortion–perception trade-off is still not well understood. This paper presents a critical review of the field spanning 1970 to 2023, drawing on 75 peerreviewed and archival sources selected through a PRISMA-informed screening protocol. We trace how image compression and quality assessment evolved together, from Shannon's rate-distortion theory through DCT- and wavelet-based standards to learned nonlinear transform coding and generative codecs. The mathematical foundations of rate–distortion theory, structural similarity, information fidelity, and the rate–distortion–perception function are then formalized, and a taxonomy is presented covering classical, structural, perceptual, no-reference, and AI-based quality metrics. Learned and generative compression architectures are examined in detail, hyperprior, autoregressive, transformer, and diffusion-based designs among them. Building on this synthesis, we propose a unified evaluation framework that folds pixel, structural, perceptual, semantic, and task-fidelity dimensions into a single adaptively weighted assessment protocol suited to modern AI-based codecs. The paper also catalogues the subjective datasets and protocols that underlie the field and presents comparative tables of representative metrics and codecs. It closes by identifying six research gaps, nine future research questions, and a short-, medium-, and long-term roadmap. The review is meant to serve as both a technical reference and an agenda-setting document for researchers, standards bodies, and practitioners working where learned compression meets perceptual quality science