Bi-Critical Reinforcement Learning Framework for Bit Rate Reduction and Quality Improvisation
G S Sandeepa, G .S Mamatha · 2024
This paper describes bit allocation for frames in H. 265 by using a bi-critical Reinforcement Learning (RL) framework. The main aim is to reduce the distortion among the Group of Pictures (GOP) under-rate and distortion constraints. Prior RL-based techniques, provide a solution only to a difficulty with limited optimization that will improve only one reward mechanism that uses rate and distortion parameters. However, these parameters are normally ad-hoc and may not give good results for various encoding circumstances and video clips. We apply the Deep Deterministic Policy Grading (DDPG-RL) algorithm, which employs two critics, to solve this problem, with one for learning for prediction and the other for rate parameters. In normal, to update the agent the distortion critic works by satisfying the rate constraint parameters. In contrast to this, rate critic always makes rate constraints with priority as and when an agent goes over a bit higher. From the experiments, In terms of rate-distortion performance, our solution surpasses the bit allocation over a base module and a single critic while providing fair rate regulation.