Research And Implementation Of Key Technologies In Video Compression

Video compression mainly uses video frame interpolation and image compression to remove inter-frame and intra-frame redundancy.However,existing video compression standards all adopt heuristic design and manual optimization.Because heuristic rules cannot fully cover the complex features of natural video,the effect of video compression cannot be guaranteed.The optimization method based on deep learning improves the effect of video compression to some extent,but there are still many problems in complex motion scenes and low bit rate states.In view of the shortcomings of existing studies on video frame interpolation and image compression,this thesis mainly carries out the following two researches:(1)In order to solve the problem that the existing video interpolation methods are difficult to estimate optical flow in complex motion scenes and lack of optical flow annotation data,this thesis presents a multi-frame video interpolation network MFVI-Net based on accurate optical flow estimation.MFVI-Net uses unsupervised training optical flow estimation network,combined with multi-scale network structure and bidirectional optical flow method to effectively expand the range of motion estimation.At the same time,MFVI-NET uses flow refine network to further refine optical flow and solve the problem of motion occlusion.Finally,an optimization method based on improved adversarial training and model pruning is presented to optimize the visual effect and time-consuming of MFVI-NET.Based on PSNR and SSIM image indexes,MFVI-NET is compared with other methods on a variety of datasets.The results show that MFVI-NET achieves better results.(2)In order to solve the problem that existing image compression methods have low image recovery accuracy and easy to lose the image texture details under low bit rate,this thesis proposed a high quality image compression method HQICM based on image sampling network.HQICM combines the image up-sampling and down-sampling network with the traditional codec(JPEG),using the image up-sampling network based on the improved adversarial training to improve the quality of image restoration at the decoder.At the same time,HQICM adopts the image down-sampling network based on texture loss to effectively retain the image texture at the encoder.Finally,to achieve high quality image compression,the strategy of joint training is adopted to solve the problem that the down-sampling network cannot be trained without labeled data and the image distortion caused by JEPG codec.Based on PSNR image index,HQICM is compared with other methods on a variety of datasets at low bit rate.The results show that HQICM achieves better results.