Coding Unit Partitions Using Depth-Wise Separable Convolution in Versatile Video Coding (VVC)
Jyotirmoy Karjee, Aryan Dubey, Anurag Chaudhary · 2023
The Versatile Video Coding (VVC) provides significant improvement in reducing the bite-rate (i.e., approx. 50 %) over High Efficiency Video Coding (HEVC), thereby reducing the coding efficiency while maintaining the same visual quality. The VVC consists of quad-tree and multi-type tree (QTMT) struc-ture where brute-force recursive search operation is executed to perform coding unit (CU) partitions. Due to the extensive recursive searching mechanism, the complexity of VVC increases drastically (thereby increasing the encoding time). To solve this issue, instead of using brute-force recursive search mechanism, we propose a Lightweight Depth-wise Convolution Neural Networks for Coding Unit Partition (LDC-CUP) mechanism to perform CU partitions in VVC. The LDC-CUP mechanism uses depth-wise separable convolutions which consists of depth-wise operation and point-wise operation can optimize the rate-distortion (RD) function. We conduct experimental validations to depict our results where the LDC-CUP mechanism provide improvements of 38.665 % for RD loss function, 26.656 % for inference (running) time of model, 35 % for binary prediction accuracy and sixty times lower bit rate computations are achieved compared to MSE-CNN VVC encoder model. Based on model inference time, we also provide justification to reduce the encoding time of LDC-CUP model compared to VVC and MSE-CNN model, respectively.