PreNet Aided DF-JSCC: An Optimized Approach for Wireless Semantic Image Transmission Over Dynamic Radio Channels

Semantic communications provide a novelly simplified and intelligent paradigm for traditional wireless image transmissions. Although the existing wireless semantic image transmission (WSIT) schemes have made some progress in combating the “cliff effect” of performance degradation, they are still insufficient in terms of adaptivity, coding flexibility, and spectral efficiency for delivering diverse image contents in dynamic wireless channels. This paper proposes a deeply flexible joint source-channel coding (DF-JSCC) scheme and a transmission performance prediction network (PreNet) to ameliorate the overall WSIT system performance. Firstly, we incorporate a multi-channel feature fusion module (MC-FFM) into DF-JSCC to enhance the capability to capture detailed image features at different scales. Then, we design a feature enhancement module (FEM) to strengthen the representation of critical features by adaptively adjusting the channel weights of the image feature maps in various radio channel conditions. Finally, we introduce the PreNet, a model that achieves high-quality PSNR prediction for reconstructed images by modeling feature distribution from detail-rich feature maps and utilizing multidimensional feature fusion. Experimental results demonstrate that, compared with DeepJSCC-V and WITT schemes, DF-JSCC achieves PSNR improvements of at least 0.93 dB and 0.44 dB, respectively, while also reducing the resource consumption by more than 50% under the same transmission conditions, which underscores its superior performance in spectral efficiency. In addition, the PreNet has enhanced the prediction quality and robustness of the WSIT in complex environments by reducing 14% of the average prediction error.