Lightweight and Adaptive Deep Coding for Wireless Image Transmission in Semantic Communication

Currently, deep learning-based joint source channel coding (JSCC) methods have achieved significant progress in enabling semantic communication. However, existing methods of this type often fall short of meeting the new demands in terms of model parameter size and storage efficiency. To address these issues, we propose a lightweight wireless image transmission method based on a tiny ConvNeXt architecture (ConvNeXt-T), referred to as DeepJSCC-T. To reduce the system’s parameter size and storage requirements, we adopt the efficient ConvNeXt-T as the backbone network for JSCC, thereby achieving a lightweight design. To further ensure adaptability to varying bandwidth and signal-to-noise ratio (SNR) conditions, we design a lightweight SNR-adaptive module based on large convolutional kernels, which enhances both flexibility and transmission performance. The experimental results show that, compared with existing methods, the proposed DeepJSCC-T achieves significant reductions in both parameter size and storage overhead. Specifically, DeepJSCC-T reduces the number of parameters by 24.65% and 27.52% relative to the Attention-based DeepJSCC (ADJSCC) and the variable-length DeepJSCC (DeepJSCC-V), while lowering storage overhead by 30.98% and 33.47%, respectively. Notably, despite its more compact architecture, DeepJSCC-T maintains competitive PSNR performance, demonstrates clear advantages in high-resolution image transmission, and exhibits strong adaptability under different SNR and bandwidth conditions, making it suitable for resource-constrained semantic communication scenarios.