Research On Multi-scale Dense Network And Its Application In Image Post-processing

The development of Internet technology and the popularization of various shooting equipment have brought more pictures to people,and the explosive growth of picture resources has put forward new challenges to image post-processing technology.The traditional image post-processing algorithm has the problems of high prior knowledge requirement and poor adaptability of the algorithm,and the use of deep learning method can automatically model and obtain better performance of the image degradation problem by learning a large number of samples,which becomes the mainstream direction of image post-processing research.The two key issues of the deep learning approach are how to design the network structure of the neural network and how to optimize the loss function.The main research tool in image post-processing research is convolution neural network(CNN),whose development trend is the deepening of the depth of the network,but less attention is paid to the extraction of multi-scale information in the image on the scale.The deepening of network depth will lead to the problem of poor network convergence,and single-scale network can not extract and utilize the multi-scale information in the picture will also limit the improvement of network performance.On the other hand,the loss function commonly used in the study of image post-processing is based on the loss function of mean square error(MSE),although the use of this loss function can bring better peak signal-to-noise ratio(PSNR),but objectively there are problems that can easily lead to image blurring.Based on the above problems,this dissertation proposes an improved method,including:Firstly,in this dissertation,a multi-scale rectangle dense network(MSRD)is proposed in the study of image post-processing,which has multi-scale characteristics and combines the idea of dense connection.Multi-scale network can effectively extract the multi-scale information in the picture to guide the repair of picture details,dense connection can improve the utilization of feature map in the network,improve the convergence of the network.Secondly,the network structure of MSRD is more complex,the number of network parameters is more,compared with other network calculation speed is slow.In this dissertation,the multi-scale loss function is set up in the training of MSRD,the different nodes in MSRD are supervised,the image output quality of different nodes in MSRD is improved,so that the network model can reduce the network level and speed up the calculation by pruning under the premise of ensuring the quality according to the different application scenarios.Finally,in this dissertation,the loss function optimization experiment is carried out for MSRD,and it is proposed that a mixed loss function based on multilayer feature can be used in the research of image post-processing.This loss function contains two components,namely,MSE loss function and feature loss function,and the performance of this loss function and the weights of two components in this loss function are verified experimentally.