Deep Learning Approach in Intra -Prediction of High Efficiency Video Coding

The basic processing unit of HEVC is CTU. It can possess various size from 64×64 to 8×8 and it increasing coding efficiency as size is large. The computational complexity is an issue to be focused as HEVC has many pros to be considered as a best video compression technique. This paper focus on reducing the computational complexity of high-efficiency video coding (HEVC) in intra prediction by using combining depth decision and deep learning techniques. The proposed method provides a neural network for depth analysis of CTU followed by a deep learning network with multiple sizes of kernels for convolution and pervasive parameters that are trainable, from the database provided. A database provided here is constructed considering both the image frame from video and encoding abilities of CU. Database has the image frame data indicating the image value of CU and a vector of 16x1 depending on CU’s encoding details. It has a label to indicate, whether the CU is split or not. Initially image frame that is of huge size is assorted to various scales and split is created. Followed by modelling the partitions into a three level classification problem. To solve classification issue, a deep learning based CNN structure that possess various size kernels and parameters for convolution is developed, that should be analyzed and learned through a database that is established. The results show a dip in the encoding time of intra mode in HEVC for the given database