Multiple Description Compressed-image Enhancement Method Based On Deep Neural Network

Different from traditional single-description coding,multiple description coding not only cause concussion artifacts in the side and center compressed images,but also cause split artifacts in the side compressed images.In this paper,we study the design of side and center compressed image enhancement network for multiple description coding as well as side and center compressed image enhancement based on a single model.We use deep learning technology to solve these problems.The main work and innovation of this paper include two aspects:1.A multiple description encoding image enhancement method combining edge and midway decoding feature learning is presented,which considers both side and center decoding image enhancement.Thus,Optimize the features of side decoding and center decoding through joint learning to achieve better network training than a single image enhancement method.First,considering the characteristics of side independent decoding and joint decoding of multiple description encoding,a network-shared side low-resolution feature extraction network is proposed to effectively extract the features of two side decoded images with the same content and different details.At the same time,a residual recursive compensation network structure is designed and used in the side and center low-resolution feature extraction network.Secondly,a multi-description side sampling reconstruction network is designed,which uses a partial network layer parameter sharing strategy,which can reduce the number of network model parameters and improve the generalization ability of the network.Finally,a multi-description road sampling reconstruction network is proposed,which fuses two low-resolution features of the side with the low-resolution features of the center to enhance the multi-description compressed image.A large number of experimental results show that the proposed method is superior to many compressed image enhancement methods such as ARCNN,Fast ARCNN,Dn CNN,WSR and DWCNN in model complexity,objective quality and visual quality evaluation.2.A multiple description compressed image enhancement method based on single modulation model learning is proposed,which combines the localization of convolutional neural networks with the global modeling advantages of Transformer to solve the problem of excessive model parameters.Firstly,a Dual-encoder network is designed to extract complementary features and increase the diversity of each scale feature,while the Dual-decoder network is designed to effectively compensate for the second decoder subnet by integrating the Dual-encoder features and the first decoder subnet features.Secondly,a Multi-Path Transformer(MPT)layer is introduced,which includes two pairs of multi-head self-attentionn modules and a Group-Ensemble Multi-Layer Perceptron(GE-MLP)module.In the proposed MPT layer,a Series-parallel Cross Fusion(SCF)mechanism is proposed to transfer the characteristics of the first pair of Transformer layers to the second pair of Transformer layers,so as to improve the non-linear expression ability of the model.At the same time,the proposed GE-MLP module first uses grouping operations to reduce the complexity of the model,and then uses combination operations to aggregate features to solve the problem of weak expression ability of a single MLP.In addition,a modulation network is designed to adjust the Dual-decoder network to achieve adaptive enhancement of side decoded image and the center decoded image.A large number of experimental results show that the proposed method outperforms many methods such as ARCNN,Fast ARCNN,Dn CNN,WSR,DWCNN,U-Net method,and Uformer.