Research On SNR Adaptive Wireless Image Deep Coding And Transmission Technology

It can be known from Shannon's separation theorem that without limiting the time delay and complexity,separately optimizing the source coding and channel coding can also achieve the best performance.Based on this theory,modern communication systems mostly use highly optimized source coding algorithms such as JPEG,JPEG2000,BPG,and near-optimal channel codes such as LDPC,Turbo codes,and polar codes to approach the theoretical optimum.In recent years,with the vigorous development of deep learning,the research of joint source channel coding in wireless communication achieves great success.However,the separate coding scheme is affected by the "cliff effect".Research on deep learning based joint source channel coding usually trains the designed network under a specific SNR,and the adaptability to the changes in the SNR is not satisfactory.Which hinders its application in real wireless scenarios,and cannot be applied to multi-user scenarios.Therefore,this paper designs and implements a SNR-adaptive coding scheme for wireless image transmission,the scheme focuses on the adaptability of the decoder to different SNRs.The scheme is based on autoencoder structure,by adding the pilot signal as the input of the decoder,the decoder gains the ability to estimate the SNR of the channel,and then adaptively decode the transmitted images.Firstly,this paper proposes a basic adaptive model and on the basis of it an enhanced adaptive model is proposed.Under different channel models and different bandwidth compression ratios,the two proposed models are trained and evaluated,and compared to the state-of-art deep learning based joint source channel coding method.Experimental results show that whether it is an adaptive basic model or an enhanced model,its signal-to-noise ratio adaptive capability is better than existing methods,and it has considerable application potential in multi-user scenarios.In addition,this paper also explores the robustness of the decoder in the proposed model when there is an error in the estimation of SNR.Experimental results show that the adaptive decoder has strong robustness to the noise in the estimation of SNR.