Wireless semantic image transmission system based on deep flexible coding strategy and performance prediction

To improve adaptivity and coding flexibility for diverse image content under complex and dynamic wireless channel conditions, a deeply flexible joint source-channel coding (DF-JSCC) scheme and a transmission performance prediction network (PreNet) for semantic image transmission (SIT) were proposed. Specifically, a multi-scale feature fusion module (MS-FFM) was incorporated into DF-JSCC to enhance detailed feature extraction. A feature enhancement module (FEM) was also designed to adaptively adjust channel weights in the feature maps to reinforce critical information. Additionally, PreNet was introduced to accurately predict the peak signal-to-noise ratio (PSNR) performance of reconstructed images by modeling feature distribution in detail-rich feature maps and utilizing multidimensional feature fusion and learning. Experimental results demonstrate that, compared with the existing state-of-the-art DeepJSCC method, DF-JSCC achieves at least 1.03 dB improvement in PSNR across various communication scenarios, while reducing bandwidth consumption by over 50% under equivalent transmission conditions. The PreNet effectively improves the prediction quality and robustness of the SIT in complex environments by reducing the mean squared error （MSE） by 20%.