Research On The Optimization Of Codec Technology For HEVC SCC

Recent proliferation of screen content video applications repid development,in-cluding screen sharing,wireless displays,distance education and so on.Video materials of these applications mainly contain computer graphics like cartoons,texts,as well as some natural contents captured by cameras.Different from the pure camera-captured contents,screen content is typically composed of some repeated patterns,massively flat areas,limited numbers of color values and without sensor noise.These characteristic-s,however,are hardly exploited by the traditional coding standards dedicated to the camera-captured videos,such as H.264/AVC,High Efficiency Video Coding(HEVC)et al.Thereby,a screen content coding(SCC)extension for the HEVC standard content videos was launched by the Joint Collaborative Team on Video Coding(JCT-VC)in January 2014.There are four major coding tools have been proposed,known as Palette mode,Intra Block Copy,Adaptive Color Transform and adaptive Motion Vector Res-olution,respectively.In this paper,we focus our research on two aspects about the optimization of SCC codec,i.e.the parallelism of palette mode and fast mode decision for All-Intra,aiming to improve the parallism of palette mode and reduce the coding complexity.The main research is as follows.As for the aspect of improving the parallism of decoder,we study the parallel optimization of palette mode.In the current palette mode,the pixel reconstruction pro-cess need to wait until all palette indices information completely parsed.Besides that,the dependence among pixels in one coding unit(CU)is very strong,which prevents the parallel reconstruction within the CU.Therefore,this paper propose a novel palette mode decoding architecture to improve the parallelism of SCC implementation,and our contributions are summarized as two points.(1)We propose only to split the largest size of CU using palette mode,since s-maller CUs do not pose any challenge to implementation.Experiments show that our algorithm can reduce the waiting time for reconstructing and in the meantime achieve a remarkable parallelism of palette mode.Compare with SCM-5.0,parallel palette mode decoding architecture can achieve an average 320%speed acceleration at the decoder with average less than 1.4%increase of BD-rate(Bjontegaard-Delta-rate),where we choose to split the largest 32x32 CU into 8 sub-CUs.(2)We conclude that the parallelism won't constantly increase with incremental number of split sub-CUs,Combining with the Rate-Distortion(RD)performance,we have also discussed about the optimal number of sub-CUs under different scenarios.As for the aspect of reducing the coding complexity,we study on fast mode de-cision under All-Intra configuration.The new coding tools significantly improve the coding performance of screen content,but also make the mode decision process more complex.During the mode decision process,encoder needs to examine every possible candidate mode for each CU to obtain the optimal mode with the minimum RD cost for the CU,and then decides whether it should be further partitioned into 4 smaller CUs for a better RD cost through the quadtree coding method.Therefore,the additional coding tools make the mode decision process more time-consuming.In this paper,we propose a content-aware fast mode decision algorithm under All-Intra configuration,which can remarkably reduce 40.01%encoding time of SCC encoder SCM-6.0 and preserve the coding performance with only 1.30%BD-rate increase.The main contributions are summarized as three points.(1)We propose a content-aware fast mode decision framework to avoid examining unnecessary modes and CU sizes.(2)We propose a Last-Bit Convolutional Neural Network(LBCNN)classifier to predict content types.Specially,we have a preprocessing operation for the input image,by extracting the last-bit of each pixel's gray value as the input instead of the gray im-age itself.It is inspired by the observation that the sensor noise of the camera-captured contents is mainly added on the last bit of the luminance component,while typical computer-generated contents don't contain any sensor noise.Benefitting by the pre-process of extracting last-bit image,the classification accuracy is greatly improved.(3)The content types of blocks predicted by LBCNN classifier is real and won't be affected by coding information,such as quantization parameters.In addition,our LBCNN classifier can be combined with rate control module to further improve the coding performance.