Research On Background Modeling Based High Efficiency Video Coding Method

With the rapid development of communication technology and multimedia technology,video media have penetrated into people’s work and life in all aspects,and become irreplaceable first media.However,the amount of video data is huge,thus the uncompressed video is almost impossible to transfer in the network,and the storage cost is also unbearable.Thus,video coding technology in the current large video data era is even more important.Video coding technology is the core technology of security monitoring,broadcast television and other applications,and video coding standards provide a unified technical specification,making video technology to be popularized.From the nineties of the]ast century,a series of video coding standards have been developed,and constantly promote the development of video technology to meet the changing needs.However,in the past few years,the explosive growth of the media,AR,VR and other new media,as well as the surveillance video with more high-definition for public security needs,are rapidly accelerating the growth of video data.The data produced in the past few years are even more than the data generated during the previous four thousand years.Hence,even the latest video standard H.265/MPEG-H HEVC has been unable to meet the real needs.Thus the new coding technology is really needed to further improve the coding performance.Background reference technology is one of the emerging technologies in video coding technology.Based on the background modeling theory,it can eliminate the redundancy of video signal by making full use of the static background characteristics,and improve the coding performance to the maximum extent.Current background picture generation models are mostly the background models used for video analysis,which require a large number of training samples,and have extensive iterative granularity.Hence,they are not suitable for video coding.Moreover,the traditional bit allocation methods used for background reference are mostly based on the empirical formula.As a result,these approaches cannot be adjusted adaptively for different content.In addition,because the intra mode cannot utilize the reference,the coding efficiency is still relatively low.Therefore,the number of required bits for intra mode is very high,which is easy to cause the transmission delay,packet loss and so on.In order to solve these problems,this dissertation focuses on the application of background modeling theory in video coding.For future(next generation)coding standard technology,this dissertation carries out research in three aspects,namely,the background reference generation algorithm,the bit allocation of inter coding for background blocks and the intra coding method for surveillance videos.The main innovations and contributions of this dissertation are listed as follows.(1)This dissertation presents an efficient and progressive background reference synthesis algorithm.A synthetic algorithm is designed for both the static camera case and the dynamic one.For the still camera sequence,the candidate background blocks are detected based on the temporal and spatial correlation of the background frame.Then,the candidates are ranked according to the spatio-temporal distribution of each background block,a limited number of background blocks are selected to be coded with high quality.Finally,these high quality background blocks are used to update the background reference incrementally.For dynamic camera video,the image alignment is conducted based on accurate global motion estimation.Then the similar algorithm in the static camera case is used to detect the background blocks,and then the illumination smoothing algorithm is added in the updating of background reference.Both of these two algorithms effectively improve the coding efficiency and avoid the bit burst due to the coding of additional background reference.In this dissertation,the progressive background reference synthesis algorithm for static camera sequences has been received by the latest video coding national standard AVS2 and integrated into the AVS2 reference software.(2)This dissertation proposes a bit allocation strategy for background reference based on stability analysis.Based on the existing bit rate allocation method,this dissertation allocates the given bit rate of the background blocks in the time domain.That is,this dissertation studies how to effectively reallocate the total given bit rate for background blocks,which could be decided by traditional bit allocation methods,between the current background blocks and succeeding collocated background blocks to achieve optimal global coding performance.By analyzing the stability of the video content,the motion distribution of each background block is extracted,and then the probability of that the current background block is referenced by subsequent blocks is estimated.Base on the estimated probability,the relationship between the coding quality of the background block in the current picture and the collocated pseud background blocks in the subsequent pictures is determined.Based on this relation,the optimal bit rate allocation scheme under the global rate distortion criterion is obtained,which can be used to guide the coding decision of background blocks.In contrast to traditional methods,the proposed bit allocation strategy for the background blocks in this dissertation not only considers the rate distortion of the current block,but also considers the impact of its distortion on subsequent blocks,which helps to achieve the optimization of global rate distortion.(3)This dissertation proposes an intra frame coding method for surveillance sequences based on illumination decomposition and deep learning.On one hand,considering that the reflection coefficients of the background part are basically the same at different moments and only the illumination changes occur,this dissertation proposes a novel intra mode for background block based on illumination decomposition.Using the illumination separation of a series of existing background frames,the reflectance coefficient map of the background frame is extracted,and stored with high quality in the location that can be accessed by any subsequent frames.Based on the high quality map,the light component of the background blocks in any subsequent random access frame can be acquired.Since the illumination signal has higher correlation than the original image,it is more suitable for intra coding.As a result,the method can obtain better coding performance and effectively reduce the number of bits required for intra coding.On the other hand,considering the existing intra prediction method cannot adjust the prediction pattern according to the content adaptively,this dissertation also proposes a novel intra prediction mode based on deep learning.In this mode,the prediction block with the original best intra mode is filled as the input block by the surrounding usable reconstruction pixels,and the output block obtained by the convolutional neural network is utilized as the prediction block in this mode.Compared with the existing intra prediction modes,the deep learning based mode makes full use of the surrounding coded information and provides a richer interpolation filtering method,which helps to achieve significant coding performance improvement.