SwinJSCC-GAN: enhanced joint source-channel coding for semantic communication with swin transformer and generative adversarial networks

The challenge of 6 G image semantic communication lies in preserving semantic integrity through effective cross-modal representation. This paper proposes SwinJSCC-GAN, a deep joint source-channel coding framework that integrates the hierarchical feature learning of the swin transformer with the perceptual optimisation of generative adversarial network (GAN). This framework significantly improves semantic fidelity and anti-interference capabilities for image transmission over Rayleigh fading channels by constructing VGG-based and PatchGAN-based deep joint source-channel coding for multi-scale semantic feature extraction and combining dynamic modulation with hybrid loss functions. Experimental results demonstrate that the proposed scheme outperforms conventional methods in terms of PSNR, MS-SSIM, and LPIPS metrics, thus verifying its effectiveness in complex wireless environments.