Reg ANR-based Image Super-resolution Algorithm Research On Its Application In Video Coding

With the rapid development of the Internet,information carriers such as images and videos,which can intuitively and vividly express information,are distributed in every corner of the Internet.People's demand for high-resolution pictures and videos is growing with the development of display terminals.However,due to the low resolution of images or videos collected in the past,or the limitations of bandwidth and storage resources,pictures and videos with insufficient resolution can not meet people's needs,while ultra-high-resolution videos and images usually occupy more memory.These situations require the use of some technology to improve the resolution of pictures and videos.Super Resolution Image Re-construction(SRIR)is a software algorithm that can convert existing Low Resolution(LR)images into High Resolution(HR)images.Compared with acquiring high-resolution images through hardware,software-based algorithms have the characteristics of low cost,simple operation and various methods.Therefore,researching on image super-resolution reconstruction and its application in video coding has good theoretical and practical application value.To solve this problem,based on the research of image super-resolution algorithm,this paper proposes an improved super-resolution algorithm based on regularization anchored neighborhood regression and integrates it into the existing HEVC standard to realize a video compression and coding framework.The main work and innovations of this thesis are as follows:(1)A super-resolution algorithm based on regularized anchored neighborhood regression is proposed.The super-resolution algorithm based on sparse representation needs to calculate the sparse coefficient during the reconstruction process,which leads to the slow reconstruction speed.Although the anchored neighborhood regression algorithm achieves the effect of rapid reconstruction through the way of neighborhood mapping,this kind of sparse dictionary-based Super-resolution algorithm usually lack the optimization of the dictionary training stage,which affects the reconstruction effect of the algorithm.To this end,this thesis proposes a super-resolution algorithm based on regularized anchored neighborhood regression,referred to as Reg ANR.In the dictionary training phase,the algorithm adds regularization constraints to the error terms of the updated dictionary atoms,so that the dictionary atoms converge quickly during training,improve the performance of sparse dictionary,and then achieve good reconstruction results.Experimental results show that,compared with other mainstream super-resolution algorithms based on sparse dictionary,Reg ANR algorithm obtains better image quality in the image reconstruction stage,and the algorithm is also very competitive in image reconstruction speed.(2)Based on the above work,a video compression algorithm based on Reg ANR is proposed.The mainstream video compression coding standards use the redundant characteristics of video data in time and space to encode the video,and ignore the large amount of a priori information existing in the video.In the case of limited bandwidth or storage,the video needs to be further compressed and the learning-based super-resolution algorithm can fully learn the a priori information in the video,and thus the original video can be down-sampled to reduce the bit rate.Therefore,this thesis innovatively proposes a video compression algorithm based on the proposed Reg ANR.The proposed method encodes and transmits the down-sampled video and dictionaries,and superresolution reconstructs the video by Reg ANR at the decoding end,thereby further greatly compressing the video.Experiment results show that: 1)after the down-sampled low-resolution video is coded and decoded by HEVC,the super-resolution reconstruction effect of the Reg ANR is far superior to that of the Bicubic algorithm;2)while maintaining the same bit rate,the reconstruction effect of the proposed video compression method is better than direct HEVC encoding.