Research And Application Of Video Image Quality Enhancement Technology Based On Deep Learning

When the transmission bandwidth or storage is limited,the compressed noise caused by video compression obviously affects the user's subjective experience.For this kind of low-quality compressed video,this paper uses the convolutional neural network to propose solutions from the spatial domain and the joint spatial-temporal domain respectively.The main research work of this paper includes:1.To improve the quality of compressed video from the perspective of super-resolution.We propose an image super-resolution model that utilizes multi-level features.The experimental results show that the proposed network model achieves an average gain of 0.22 dB compared with the very deep super-resolution network VDSR on general four test sets.2.Using intra-frame pixel correlation in spatial domain to improve the quality of compressed video.Using local and global residuals,a single frame network structure based on residuals block is designed.The results show that the proposed network model achieves an average performance gain of 0.22 dB relative to H.265/HEVC.3.Joint spatial-temporal domain information improves the quality of compressed video.Combining the temporal domain information of video,two kinds of multi-frame quality enhancement methods are proposed,one is LMVE based on optical flow information,the other is STMVE based on prediction.Experimental results show that the proposed LMVE obtains an average performance gain of 0.38 dB relative to H.265/HEVC.The proposed STMVE obtains an average performance gain of 0.39 dB.4.A quality enhancement method for traffic surveillance video is designed.First,the YOLO network is improved to detect the grouping vehicles within a frame,thus generating an effective training set.Secondly,a Residual Squeeze-and-Excitation Network(RSE-Net)is designed to achieve good nonlinear mapping using only 200,000 parameters.Experimental results show that compared with H.265/HEVC,our method achieves an average PSNR gain of 0.29 dB,and produces visually pleasing results when applied to compressed surveillance video.