Research On Multiple Description Coding Based On Convolutional Neural Networks

In recent years,with the development of computer networks,images and videos have become more and more widely used in networks,and their transmission has become more and more important.How to ensure the high efficiency transmission of images and videos on the channel has become one of the hot topics of research.The rapid growth of multimedia services on lossy channels has promoted the development of efficient,reliable and adaptable coding technology,multiple description coding(MDC)has become one of the most effective coding technologies to solve this situation,especially in the case of unreliable data transmission in practical applications.In multiple description coding,the source is encoded into multiple descriptions containing controlled redundancy,these redundancy are used to solve the problem of unpredictable packet loss during cross-channel transmission,the description is transmitted to the decoder through different network paths,the decoder uses the received description to reconstruct the source data.In addition,convolutional neural networks(CNNs)are widely used in the fields of image classification and pattern recognition,and they are increasingly applied to the field of image compression,and use their own advantages to obtain good visual results.This thesis starts with improving the accuracy and coding efficiency of the reconstructed images,and uses the knowledge of deep learning to conduct an in-depth research on multiple description coding methods based on convolutional neural networks,and then studies multiple description images coding method based on convolutional neural networks,the main innovative research results are as follows:1)A multiple description image coding method based on symmetric convolutional auto-encoder(SCAE)is proposed.The symmetric convolutional auto-encoder is used to set up a full convolutional network,uses a convolution layer with a stride of 2 instead of the pooling layer,and alternately connects it with the convolutional layer to extract image features layer by layer,and sets deconvolutional layer maps low dimensional features into high dimensional inputs.The encoder of the multiple description coding framework encodes the extracted convolutional features into two descriptions containing two subsets,which are transmitted to the decoder through different network paths,and the probability of all descriptions being lost at the same time is very low,thus avoiding the problem of packet loss to the greatest extent.The experimental results prove that the scheme can solve the problem of image quality degradation caused by packet loss,and improve the accuracy of reconstructed image while ensuring the efficiency of image coding.2)A multiple description coding network(MDCN)based on convolutional auto-encoder is proposed.The network framework is built on the basis of convolutional auto-encoder,in order to achieve high quality image compression at low bit rates,the multiple description encoder network and multiple description decoder network are seamlessly integrated into the end to end compression framework.Because the rounding function of quantization is not differentiable,this thesis uses additive uniform noise to imitate quantization noise in the optimization process.By combining SSIM loss and distance loss to train the multiple description encoder network,it is ensured that structural information can be shared even if the features are divided into multiple descriptions.The experimental results prove that the performance of the proposed scheme is better than other existing schemes,which can reduce the distortion of the reconstructed image,especially at low bit rate,the quality of the reconstructed image obtained is greatly improved.