LeNet5-Based approach for fast intra coding
Amna Maraoui, Imen Werda, Sayadi Fatma Ezahra · 2021
Recently, High Efficiency Video Coding (HEVC) standard has emerged to achieve high coding efficiency performance when introducing several specific modules. One of the most influential new tools is code the tree unit (CTU) partition structure. In the HEVC standard, a CTU becomes a block-shaped region replacing a macroblock based on H.264/AVC. This module achieves a significant gap of compression performance at the expense of encoding complexity increases under All-intra configuration. Block unit structure can be split into nested coding units (CUs) (range sizes 64×64 to 8×8) after applying the Rate-Distortion Optimization (RDO) process which becomes a bottleneck for feasible implementation. To resolve this issue, we suggest substituting the brute force RDO search with a LeNet-5 model for a CU partition at intra-mode. First, we created a database for the HEVC intra-mode in order to learn the learning model. Subsequently, a LeNet-5 model is proposed for predicting the HEVC CU partition. Experimental results indicate that the proposed algorithm can speed up the CU partition structure by reducing the intra-mode encoding time by an average of 87.47% while degrading by 3.64% the compression performance. Our CU decision scheme outperforms all state-of-the-art solutions in terms of complexity reductions.