A New Deep Joint Source Channel Coding with GAN Discriminator for Wireless Image Transmission
Yan Xu, Linjiang Shen, Shuqing Qiu, Chao Cui, Jundong Xu · 2024
This paper proposes a novel approach to wireless image transmission by combining Deep Joint Source Channel Coding (DJSCC) with a Generative Adversarial Network (GAN) discriminator. Our method is capable of significantly improving data compression and transmission efficiency, particularly in challenging wireless environments with limited bandwidth and high noise levels. A key innovation lies in integrating the discriminative mechanism of GANs, which learns implicit image features, with DJSCC. This fusion enables semantic-aware compression and denoising. Additionally, a double gradient update mechanism is employed during training, which enhances the fidelity of the decoded image by improving both pixel accuracy and structural similarity. Extensive experiments demonstrate that this new model significantly improves upon traditional DJSCC models in image transmission quality, particularly under low signal-to-noise ratios. The model effectively enhances structural features and visual perception quality, making it suitable for practical wireless communication applications.