Research On Post-process Baesd On End-cloud Collaboration In Surveillance Video Compression System

This paper forcuses on the compression processing method for surveillance video.For video compression processing,there are so many technical methods for general video coding,but processing technologies for surveillance video are relatively rare.Therefor,this article proposes a compression processing method for surveillance video.With the Hi3559AV100 development board,the video encoding and decoding work is realized.Subsequently,in order to further improve the image quality,we implemented the image quality enhancement task of reconstructed video.In the image quality enhancement task,we try to collect surveillance video and reconstruct it through the Hi3559AV100 development board to make a data set for training,verification and testing.Taking into account the particularity of the codec video of the development board,we have separately constructed models for processing I frame and P frame.When constructing the network for I frame,we considered the loss caused by intra-frame prediction and aimed at it.And,the distortion phenomenon is imporved.When constructing a network model for Pframe,we additionally considered the loss caused by inter-frame prediction,and imporved the quality of the images.Based on the existing sub-compression processing methods for surveillance video,in order to further compress the bit rate of the video,we tried to downsample the video before encoding,and perform super-resolution enhancement on the video after decoding to restore the original video.We still made a data set for super-resolution enhancement based on the collected surveillance video.In terms of methods,we proposed a single image superresolution method and a reference-conditioned super-resolution.In the referenceconditioned super-resolution method,we divide the video into odd and even frames,reconstruct the odd frames directly,and perform the down-sampling operation on the even frames and then reconstruct them through the development board.The obtained even frame uses the previous frame as a reference frame for super-resolution enhancement operation.The model adds an attention mechanism to use the texture information in the reference frame image to restore highresolution images,and the even frames after the super-resolution are combined with the odd frames to synthesize the output video.The single image superresolution method uses sub-pixel convolutional layers to improve the resolution,and achieve image reconstruction with multi-layer convolutional layers.The proposed compression method for monitoring video can not only ensure the quality of video image,but also improve the compression ratio effectively.In order to further improve the processing efficiency,we further realized the embedding of super-resolution enhancement technology and image quality enhancement technology into the Atlas 200 DK development board,which greatly improved the processing efficiency and achieved certain results.