Research On Video Bit-depth Enhancement Based On Convolutional Neural Networks

With the development of information technology,people’s requirements on the visual quality of multimedia resources are increasing day by day.In recent years,high-bit-depth display devices have become increasingly developed.However,the mainstream multimedia resources,especially video sources,are still at relatively low bit-depth.False contour and color distortion problems are prone to be caused when these low-bit-depth resources are displayed on high-bit-depth display devices.In order to alleviate this problem,some researchers propose image-based bit-depth enhancement algorithms based on interpolation methods and convolutional neural networks.However,to our knowledge,there is no bit-depth enhancement method designed for videos.And when these single-image-based approaches are directly applied to video-processing tasks,the information between adjacent frames will be ignored and inter-frame continuity can’t be guaranteed.Therefore,with the utilization of convolutional neural networks,this paper proposes two bit-depth-enhancement algorithms specially designed for videos as follows: 1.A spatiotemporal symmetric convolutional neural network for video bit-depth enhancement.In this paper,we first proposed a symmetric encoder-decoder network.Input continuous frames at low bit depth are first motion compensated and then put into sub-networks in the encoder respectively.And convolution filters in temporal symmetric sub-networks share weights to achieve a lower model complexity.In addition,symmetric skip connections are introduced between spatial symmetric encoder and decoder,to transfer features and alleviate gradient diffusion problem.Experimental results show that the symmetric network structure can reconstruct high-bit-depth video frames with good subjective.Compared with the image bit-depth enhancement algorithms before,this algorithm has improved the peak signal to noise ratio by approximately 2d B.2.A content-continuity-guided multi-frame bit-depth enhancement model based on convolutional neural networks.Considering the requirement of real-time performance and inter-frame continuity of video processing problems,we remove the motion compensation module and sub-networks,only retaining the spatial symmetry structure.Besides,Gram matrix is introduced in representation of the continuity and correlation among consecutive frames.And we use the mean square error between the Gram metrics of output high-bit-depth frames and the ground truth as the inter-frame continuity loss.During the training procedure,content loss and continuity loss guide network to realize real-time end-to-end multi-frame bit-depth enhancement.Compared with the algorithm proposed in scheme 1,not only the running time is shortened to 0.14%,but also the performance is improved to some extent.