Fast Intra Coding Algorithm For Screen Content Based On Convolutional Neural Network

Screen Content Coding(SCC)is an extension of High Efficiency Video Coding(HEVC),and adopts 35 kinds of intra modes of HEVC and two new coding modes: Intra Block Copy(IBC)mode and Palette(PLT)mode to improve the coding efficiency of screen content video.However,the exhaustive search for the optimal mode among these three types of mode candidates and the flexible quadtree-based coding tree unit(CTU)division structure bring a significant computational burden to the SCC encoder.For applications with limited computing power,reducing the coding complexity of SCC is very important.Therefore,how to effectively reduce the computational complexity of SCC intra coding is the research content of this thesis.For the coding unit(CU)division process of SCC intra coding,a fast algorithm for intra frame CTU depth range prediction based on convolutional neural network(CNN)is designed in the thesis.In order to predict the depth range of CTU,a CNN structure suitable for CU partition characteristics is designed in the thesis.And In this thesis,the classification label is set according to the statistical analysis of CTU depth range.The designed CNN structure is divided into three layers and uses convolution kernels of different sizes to extract features related to CTU depth.By simultaneously convoluting on multiple scales to extract features of different scales,the features are more abundant and the accuracy of classification judgment is improved.During encoding,the CTU depth range is predicted by calling the trained CNN model.By skipping and terminating the calculation of the rate distortion generation value in some depth,the coding complexity is reduced.After implementing this algorithm on SCM8.0,the encoding time was reduced by 48.34% on average.Although the above algorithm has achieved good results,the algorithm only considers the CU partition and does not consider the Prediction unit(PU)mode selection.Therefore,a fast algorithm of CNN based CU partition and PU mode selection is proposed in this thesis.First,the CNN structure for CU division and PU mode selection is designed,and the modes are divided into six categories according to the PU mode proportion analysis.The proposed structure takes the entire CTU as input and outputs the prediction labels of 85 CUs in four branches.In addition,in order to improve the accuracy of prediction,a dynamic adaptive threshold Thx is proposed,which is determined by the optimal prediction mode of neighboring CUs.During encoding,by calling the trained CNN model and combining with the threshold Thx,the six types of patterns are checked for further judgment.If all modes are judged not to be checked,the CU needs to be divided down.Experimental results show that,compared with SCM8.0,the encoding time is reduced by an average of 43.23%.Compared with the fast algorithm of intra CTU depth range prediction based on CNN based screen content coding proposed in this thesis,this algorithm can reduce the coding loss while ensuring that the coding time is greatly reduced.