Deep Joint Source-Channel Coding for Wireless Image Transmission: One-Model-to-Many

In recent years, researches on deep joint source-channel coding (DeepJSCC) for wireless communications achieve great success owing to the employment of deep learning (DL). However, all existing DeepJSCC schemes only work with a single bandwidth, ignoring the fact that different users have different bandwidths. Under additive Gaussian white noise (AWGN) channel, we present a DeepJSCC scheme utilizing auto-regressive model and transformer block for wireless image transmission, where a single model is trained across multiple compression ratios (CRs) through masking, enabling it to adapt to various bandwidth CRs. Specifically, at the transmitter, our encoder first combines convolutional kernels with transformer block to effectively capture both global and local features of an image. Then we transmit a portion of the encoded image by masking. At the receiver, based on the masking strategy, we use auto-regressive model that fully utilizes the known prior information to restore the masked portions of the encoded image. Extensive experiments show that our scheme achieves the adaptation of a single model to various CRs and better image reconstruction quality than the separately models trained under fixed CRs.