We investigated methods to reduce transmission bandwidth in video streaming environments while maintaining equivalent average subjective quality. One such method, Adaptive Quantization (AQ), computes optimal block-level QPs based on perceptual metric and Rate-Distortion (R-D) principle. In Bichon [1], an AQ algorithm was proposed that computes block-level QPs optimal from a global R-D perspective for a given perceptual metric$(\Psi)$. Their approach approximated the global R-D optimization problem using the high-bitrate approximation, the R-D Shannon lower bound, and the rate independency assumption. These approximations enabled theoretical derivation of block-level delta QPs to maintain a target GOP bitrate. However, inaccuracies arose due to (1) low correlation between the Shannon bound and real-world R-D relationships and (2) rate dependencies caused by reference QP changes. These limitations led to significant deviations between the actual encoding results and the target bitrate. To address these challenges, we propose an improved AQ method that replaces the high-bitrate approximation and Shannon bound with the linear relationship between$log(\lambda)$and QP, and between SATD and bitrate R.

This paper proposes a novel video encoding framework that combines Evolution Strategy (ES) [1] and a Vision Transformer (ViT) [2] to optimize Quantization Parameters (QPs) for improved rate-distortion performance in Constant Bitrate (CBR) encoding using the x265 video encoder. The framework consists of two key components: