Study On Fast Coding Unit Partition Algorithm Based On HEVC

With the emergence of high resolution digital video,the amount of data has risen dramatically,which brings great difficulties and challenges to real-time transmission and storage.Therefore,video compression is very important to us.Compared to the previous generation coding standard,the compression efficiency of High Efficiency Video Coding(HEVC)is doubled under the same visual effect.Due to the adoption of many new technologies,the coding complexity is greatly improved.Among many new technologies,the partition of coding units(CUs)based on a quadtree structure is of vital significance for complexity reduction.In recent years,traditional methods rely on the intermediate features to simplify the process of Rate Distortion Optimization(RDO).However,the intermediate features are complex and insufficient.Futhermore,these features are difficult to find for our research.In order to solve these shortcomings,this paper has designed the following two solutions.The optimization algorithm is designed based on spatial correlation.The current CU must be similar or related to its neighbor CUs in texture.In addition,the Z-scan encoding style makes sure that the encoding of left and upper CUs had been completed when the current CU is under encoding,which can be used as a reference.Based on the above two premise,this paper calculates the regulation between current CU and its surrounding CUs in the sample,and proposes reasonable and complete assumptions for the various possibilities in the process of determining the coded depth range.Following the top-down rate-distortion cost comparison process in the original compression scheme,a fast CU partitioning algorithm based on spatial correlation is designed.By skipping or prematurely terminating some CU partitioning processes,the complexity is reduced and the encoding process is accelerated.Convolution Neural Nettwork(CNN)performs well in many computer vision tasks such as image classification.CU partition can be regarded as a combination of binary classification problems.Therefore,CNN is also suitable for predicting CU partitioning.In this paper,a large enough data set is integrated,and a convolutional neural network including three convolution layers is designed.The three-level CU partitioning feature is extracted,and the coding depth of each CU is directly obtained.The encoding speed can be improved because the traversal process of the top-down coded depth full search is avoided.In this paper,HM16.5 is used as a benchmark.By compressing 18 test video sequences of different resolutions under four Quantification Parameters(QP),this paper get the baseline of subsequent work.According to the above two optimization algorithms,experiments were carried out for the outcom of bitrate and encoding time.Compared with the HM16.5 benchmark value,the algorithm are verified in terms of subjective visual effects and objective evaluation criteria.The experimental results show that the encoding time can be effectively reduced with negligible bitrate loss.This paper also analyzes a series of accidental results in the experiment and looks forward to the direction of the research.