Reducing HEVC Inter-Prediction Complexity: An Attention-Based Network
Abstract
As an advanced video coding standard, High Efficiency Video Coding (HEVC) can remarkably reduce the bit-rates for equal perceptual video quality. Compared with H.264, the great promotion partly comes from the usage of the deeper coding quad-tree. The deeper coding quadtree provides HEVC the ability to encode higher resolution video and reduce more bit-rates, however, it significantly increases the time complexity at the meanwhile, since HEVC searches for best coding tree depth by traverse every node exhaustively. Therefore, this thesis proposes a network to bypass the brute-force traversal and obtain the eventual coding unit (CU) partitions. First, a database, which contains each frame and corresponding CUs generated from 111 videos, established by running HEVC official software HM. Then, feed frames into a proposed convolutional neural network (CNN) in [6] to extract features. Third, the extracted features are fed into our proposed attention-based compression (ABC) network. The output of the attention-based network is predicted CUs. Finally, the experimental results shows that we build a trade-off between prediction accuracy rate and running time and can reduce the computational complexity significantly for HEVC.
Citation
Chen, Tianrui (2021). Reducing HEVC Inter-Prediction Complexity: An Attention-Based Network. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /196110.