VONV: NR-VMAF Based Optimization for VSR Quality Assurance and Thermal Throttling Mitigation

초록

On-Device VSR(Video Super Resolution) is challenged by the high computational load of Deep Neural Networks(DNNs), causing heat and performance degradation. Adaptive VSR offers a solution by optimizing performance, but it must also take a balanced approach that considers not only maintaining performance but also ensuring visual quality. This paper proposes VONV(VSR Optimization with NR-VMAF), an adaptive VSR method that uses NR(No-Reference)-VMAF to selectively apply DNN-based SR or bicubic interpolation per frame, in order to mitigate thermal throttling while maintaining consistent video quality. NR-VMAF demonstrates high accuracy, staying within 2 points of FR-VMAF, validating its reliability. Tests on two devices with distinct performance profiles showed VONV extends the thermal throttling threshold threefold and reduces overall inference time by 1.4 to 4 times. These results demonstrate that VONV, by using NR-VMAF, effectively optimizes both quality and performance in mobile environments. © 2025, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드