Low-complexity CNN-based CU partitioning for intra frames

The High-Efficiency Video Coding (HEVC) standard has high compression efficiency. This efficiency is achieved at the expense of increasing the computational complexity. The HEVC encoder has the hierarchical search for optimal Coding Unit (CU) partitioning. It is based on rate-distortion optimization. Various solutions are proposed to reduce the encoding time. But, the machine learning-based methods have more effective in reducing the encoding time. Yet, deep learning tools have a relatively high computational load. So, in this paper a new low complexity convolutional neural network has been designed. It is called Convolutional Neural Network-based CTU Partitioner (CNNCP). It reduces the computational complexity of the HEVC encoding. The CNNCP takes the CTU luminance component and the quantization parameter (QP) as inputs, and provides the CU depth matrix in output at once. The CNNCP does not follow the hierarchical approach. Thus, it has a fixed computation structure that facilitates the use of parallel processing tools. The CNNCP has a simple structure with a least number of parameters, and thus, it has the least computational complexity. It has been trained and tested with a large database for all QP values. The results show that it reduced the encoding time by more than 90%, and makes it suitable for real-time applications.