A Fast CTU Depth Selection Algorithm for H.265/HEVC Based on Machine Learning

This paper presents a fast depth selection algorithm for CTU (frame coding units) based on machine learning. In view of the fast depth selection algorithm for CTU based on machine learning, due to the lack of the depth discrimination in the initial division of coding units and the inefficiencies of the coding efficiency caused by the input feature selection of the classifier, The paper firstly design the initial division depth prediction strategy based on the texture complexity and quantization parameters to skip some nonessential sizes of coding unit by analyzing the relationship between the texture complexity of the coding unit, the quantization parameters of encoder and the depth selection of the coding unit, and by combining the texture complexity and the quantization parameters to predict the initial dividing depth of the current coding unit. Secondly, by exploring the relationship between the bit-rate, distortion and the depth selection of the coding unit, the input characteristics of the classifier are determined and the selection strategy of the coding unit termination depth based on the bit rate and distortion is designed. Finally, the partition problem of the coding unit is modeled as the problem of the two-element classification and the nearest neighbor classifier is used. By skipping the calculation process of the time-consuming rate distortion cost, the ending dividing depth of the current coding unit can be judged in advance and accelerate the process of the inter-frame coding. Experimental results show that the proposed algorithm can decrease the 34.56% of the frame encoding time, while maintaining the accuracy of the coding unit compared with HM-15.0.