pVAR: An Ultra Low Complexity Image Quality Assessment via Spatial Error Variance

作者：A. Dziembowski · 发表于：IEEE Signal Processing Letters · 年份：2026 · DOI：10.1109/lsp.2026.3708417 · 研究领域：Computer Science

While complex metrics, such as the structural similarity index, neural-network-based, and multi-scale approaches, offer high correlation with human perception, their computational complexity and floating-point-based implementations often limit their use in resource-constrained and real-time environments, such as edge devices and ultra-low-latency video coding. Conversely, standard metrics like PSNR are computationally inexpensive but fail to accurately model the Human Visual System (HVS), particularly in the presence of luminance shifts and non-linear distortions. In this paper, I propose a novel, hardware-friendly full-reference Image Quality Assessment (IQA) metric – Perceptual Variance (pVAR) – based on error variance and psychophysical saturation. By isolating the AC component of the error signal, the proposed method effectively simulates human contrast masking. Crucially, the core spatial aggregation relies entirely on deterministic, integer-based operations, avoiding floating-point rounding errors and enabling massive parallelization via single instruction, multiple data (SIMD) vectorization. The non-linear quality mapping is applied only at the final stage. Extensive evaluations across diverse IQA databases (TID2013, KADID-10 k, SJTU-4 K, CVIQ, CSIQ, and LIVE-IQA) demonstrate that the proposed pVAR metric achieves an average PLCC of 0.783, substantially outperforming both PSNR (0.552) and SSIM (0.664), thereby offering an optimal trade-off between perceptual accuracy a...