Multi-Network Joint Source-Channel Coding Over Varying Noise Channels

Inspired by the recent success of deep learning methods for joint source and channel coding (JSCC), we propose an optimized system to improve the performance in the transmission of deep-JSCC in an environment of time-varying noise. The scheme incorporates three aspects: First, a multi-network parallel structure is proposed to sufficiently exploit the label information. Second, a closed-form linear encoder \& decoder pair is adopted at the input and output end of the channel to handle the noise variation under the power constraint. The linear encoder and decoder update themselves in terms of the estimation on the channel noise from the pilot signals. Third, the estimation of the noise statistics is fulfilled by a transfer learning algorithm, which outperforms the usual estimation methods when the noise statistics are time-dependent. These three ingredients are integrated efficiently together in the scheme to form a complete image transmission system. Numerical results show that our optimized scheme performs better than the existing schemes in the literature.