SPAIK: нейромережева оптимізація бітрейту відео для мобільних пристроїв через super-resolution та перцептуальні втрати

Abstract

This thesis focuses on neural-assisted bitrate reduction for mobile viewers in SPAIK (Streaming Platform for Adaptive Interactive Communications). The main problem is to decrease video bitrate without perceptible quality degradation on small screens, where classical encoder optimizations often trade sharpness for compression. We propose a practical pipeline that combines modern codec efficiency (AV1) with learning-based super-resolution (SR) on the client side. The approach transmits a lower-resolution stream and reconstructs perceived detail using SR models trained with perceptual losses, which are known to correlate better with human perception than pixel-wise losses. The work outlines microservice integration for encoding and optional neural enhancement, targeted at a TypeScript/Golang backend and multi platform clients. Evaluation is planned using a perceptual video quality metric (VMAF) to quantify the bitrate–quality trade-off and to guide tuning decisions