Lightweight Generative Joint Source-Channel Coding for Semantic Image Transmission with Compressed Conditional GANs.

Dongshan Ye,Xijun Wang,Xiang Chen

DOI: https://doi.org/10.1109/icccworkshops57813.2023.10233814

2023-01-01

Abstract:Deep learning-based Joint Source-Channel Coding (JSCC) has shown promising results in wireless image transmission by integrating deep generative models. However, the computational demands of these models pose a challenge for practical deployment. In this paper, we propose a novel approach called Lightweight conditional GAN-based Deep Joint Source-Channel Coding (LGJSCC) to address this challenge. LGJSCC introduces a compressed generator using a GAN compression method based on intermediate feature distillation, enabling efficient utilization of computational resources at the receiver side. To enhance the perceptual quality of the reconstructed images, we incorporate the Learned Perceptual Image Patch Similarity (LPIPS) loss along with pixel-wise distortion, adversarial loss, and distillation loss. Experimental results demonstrate that the compressed generator achieves a size reduction of only 4.45% of the original, with Multiply-Accumulate Operations (MAC) decreasing to 11.92% of the original, while maintaining comparable image reconstruction quality. This makes LGJSCC highly suitable for deployment on resource-constrained devices such as mobile and IoT devices.