Reducing Time for Rate-Distortion Optimization in HEVC Using Attention Mechanism

Abstract

High-Efficient Video Coding(HEVC) is the video coding standard proposed by the ITU-T Video Coding Experts Group and the ISO/IEC Moving Picture Experts Group. HEVC, also termed H.265, can support 8K UHD videos and saves upto 50% bit-rate saving over its predecessor, Advanced Video Coding(AVC/H.264) standard. HEVC similar to AVC uses a recursive quad-tree-based algorithm on its Coding Tree Unit(CTU) to partition it into multiple Coding blocks(CB) for its encoding, where it uses a brute-force approach to find the best possible partition of CTU as a part of the Rate-Distortion Optimization(RDO). This part of the RDO search is very computation-ally expensive and time-consuming. This work proposes a deep learning-based approach using the CNN and the attention-based network to reduce the time to find the best partition of CTUs. A large database of videos is used to generate the inter-mode CTU partition data using the HEVC soft-ware for the training. CNN is used to extract the features within the frame and then the attention based network uses the temporal correlation of frames to improve the performance of the network. We found that there was a reduction of about 40% of the time taken to encode, with a minute loss of less than 1% drop in bit-rate savings for any PSNR and supports all the configurations, outperforming the proposed state-of-the-art approaches to reduce the complexity of HEVC.