Learning Joint Source-Channel Coding for Wireless Image Transmission: A Benchmark

Deep learning-based joint source-channel coding (JSCC) has demonstrated remarkable advancements in wireless image transmission. However, most existing evaluations assume stationary channel models, such as additive white Gaussian noise (AWGN) and Rayleigh fading channels, which do not capture the features of practical wireless environments where separate source-channel coding (SSCC) is already deployed (e.g., in 4 G and 5 G systems). Furthermore, the lack of a unified and fair benchmark has prevented researchers from conducting objective comparisons among JSCC schemes. To address these challenges, we introduce JSCCBench, an end-to-end JSCC benchmark that extends JSCC evaluation by incorporating orthogonal frequency-division multiplexing (OFDM) and a diverse set of non-stationary wireless channel scenarios. Experimental results indicate that current JSCC schemes possess a certain degree of robustness and effectiveness in more realistic wireless communication scenarios. However, extensive results also reveal the shortcomings of the schemes in some wireless communication scenarios. Specifically, we evaluate several representative JSCC approaches. Even those with channel-adaptive designs show significant room for improvement, while others achieve only moderate performance. The JSCCBench framework is publicly available at https://github.com/tianjian00/JSCCBench, providing a standardized platform for future JSCC research and development.