Multi-description Image Coding Method And Research Based On Deep Learning

With the development of Internet technology,Multiple description coding(MDC),as an anti-error transmission technique to solve the problem of packet loss or false code distortion during the transmission of images in the channel,can effectively improve the robustness of data transmission.In recent years,due to the outstanding performance of deep learning methods in the field of computer vision.Therefore,this paper combines deep learning with multi-description image coding methods to improve the quality of edge path image reconstruction and center path image reconstruction with small size image and large size image,respectively,and the main research work of this paper is as follows.(1)A multi-description coding method based on convolutional neural network for small-size images is proposed.To address the shortcomings of small-size images in reconstruction performance,firstly,a model of multi-description image coding is established by using convolutional autoencoder,which includes two parts of multiple description convolution encoder network(MDCEN)and multiple description convolution decoder network(MDDEN),and the overall performance of the network is improved by end-to-end joint training.Secondly,it is proposed to optimize the network structure by replacing the quantization noise with additive uniform noise to solve the problem of non-conductivity of the traditional quantization function.The experimental results show that the method outperforms other existing multi-description coding methods in terms of performance for both edge-road reconstructed images and center-road reconstructed images.(2)A multi-description coding method based on an improved U-net network is proposed.The method is used to improve the image reconstruction quality with the advantage of U-net network combined with image context information and fast training speed.Firstly,the model framework consists of two parts: U-net coding network and MD decoding network.The images are pre-processed and input to the U-net coding network to obtain two descriptions with differences by down-sampling and upsampling in the convolutional layer.Subsequently,it enters the MD decoding end for the image reconstruction task to obtain the two-sided decoded image and the centerway decoded image.After experimental verification,the method is improved in comparison with other literature methods in terms of PSNR and SSIM.(3)A multi-description coding method based on SENet+U-net network is proposed.The method is mainly based on improving the multi-description coding method of the U-net network and introducing the SENet algorithm,which enables the network to learn the feature information of important channels and suppress the unimportant feature information to improve the image reconstruction quality.Finally,the model is compared by establishing different bit rates.The experimental results show that the proposed method has significantly improved the subjective and objective performance evaluation indexes on the quality of multi-description image reconstruction.