Model-Driven Deep Joint Source-Channel Coding over Time-Varying Channels

Abstract:
Machine learning techniques have recently offered powerful tools for designing joint source-channel coding (JSCC) schemes with impressive performance results. Recent works have shown that deep learning (DL) based JSCC schemes (Deep-JSCC) for wireless image transmission can outperform stateof-the-art separate source and channel coding designs. However, existing works on DL based JSCC schemes either require many models trained for specific channel conditions, increasing the storage requirements, or need to introduce additional parameters in the model to adapt to new channel conditions, increasing the complexity of the architecture. In this paper, we propose DeepJSCC-MMSE, an end-to-end optimized JSCC scheme for wireless image transmission over wireless fading channels, in which a single DeepJSCC encoder/decoder pair is trained in concatenation with a model-driven channel equalization module at the receiver, which can adapt to any channel condition without additional model parameters. We show that DeepJSC-MMSE can outperform parameterized equalization techniques, without requiring the encoder and decoder to be retrained for new channel conditions, paving the way for generalized DL based JSCC schemes.