Paper accepted @ QoMEX 2026

Title: Asymmetry-Aware No-Reference Video Quality Assessment via Dual-Region Temporal Modeling

Authors: Yeganeh Chatri, Hadi Amirpour

Abstract: Modern content-adaptive video encoding increasingly relies on asymmetric compression, where semantically important regions are preserved at higher quality than background areas. This results in spatially and temporally heterogeneous distortion patterns that challenge conventional no-reference video quality assessment (NR-VQA) models, which typically assume spatial homogeneity.

In this work, we propose a lightweight dual-region NR-VQA framework that explicitly models distortion heterogeneity by jointly analyzing global context and a content-focused region using a shared ResNet-18 backbone with temporal mean aggregation. To address limited training data, a two-stage freeze–unfreeze optimization strategy is employed for stable learning.

Experiments on the QoMEX Grand Challenge dataset show that the proposed method achieves an SROCC of 0.881, the highest among the evaluated NR-VQA baselines in our experiments, including NIQE, BRISQUE, DOVER, and Q-Align. Additional evaluations on KoNViD-1k and LIVE-VQC indicate consistent generalization across datasets. These results highlight that explicit modeling of spatial heterogeneity is an effective and practical design principle for NR-VQA under asymmetric compression scenarios.