Video Compression Based On Deep Learning

With the advent of the multimedia era,lots of video is present in people's daily lives.Video has the characteristics of strong continuity and large data volume,so video compres-sion is important for storage and transmission.In recent years,deep learning has achieved breakthrough in many fields,and new solutions for video compression technology have also been given.Video compression based on deep learning has become a new field that has attracted much attention.Although research in this field has achieved certain results at present,there is still room for improvement in video compression performance,which is worth continuous research.This thesis investigates video compression based on deep learning,mainly including intra-frame compression and inter-frame compression.The specific research is as follows:(1)In order to improve the intra-frame compression and inter-frame compression of the convolutional neural network,inspired by the attention mechanism,this thesis designs a dense block attention module.The attention mechanism is introduced into the convolu-tional neural network to improve the its fitting ability.In addition,inspired by the dense connection network,the dense block attention module adopts a dense connection design,so that the module can be fully trained.(2)This thesis proposes an intra-frame compression algorithm based on attention mechanism.In order to achieve intra-frame video compression,the algorithm uses a con-volutional neural network to extract feature from the input image,quantizes the feature and arithmetic coding.By introducing the attention mechanism and quality enhancement network,the intra-frame compression performance of the algorithm is significantly im-proved.Compared with the existing intra-frame compression algorithm based on convo-lutional neural network,the algorithm in this thesis has obvious advantages in compression performance.(3)This thesis proposes an inter-frame compression algorithm based on multi-frame information fusion.The algorithm uses optical flow for motion compensation,eliminating the time redundancy of the video.In order to reduce the loss caused by video compres-sion,a reconstructed frame quality enhancement network is designed,which uses multiple reference frame information and a convolutional neural network to enhance the quality of the reconstructed frame.By introducing the attention mechanism and reconstructed frame quality enhancement network,efficient inter-frame compression is realized.