Adaptive Deep Joint Source-Channel Coding for One-to-Many Wireless Image Transmission

Deep learning based joint source-channel coding (DJSCC) has recently made significant progress and emerged as a potential solution for future wireless communication. However, there are still several crucial issues that necessitate further in-depth exploration to enhance the efficiency of DJSCC, such as channel quality adaptability, bandwidth adaptability, and the delicate balance between efficiency and complexity. This work proposes an adaptive deep joint source-channel coding scheme tailored for one-to-many wireless image transmission scenarios (ADMIT). First, to effectively improve transmission performance, neighboring attention is introduced as the backbone for the proposed ADMIT method. Second, a channel quality adaptive module (CQAM) is designed based on multi-scale feature fusion, which seamlessly adapts to fluctuating channel conditions across a wide range of channel signal-to-noise ratios (CSNRs). Third, to be precisely tailored to different bandwidth resources, the channel gained adaptive module (CGAM) dynamically adjusts the significance of individual channels within the latent space, which ensures seamless varying bandwidth accommodation with a single model through bandwidth adaptation and symbol completion. Additionally, to mitigate the imbalance of loss across multiple bandwidth ratios during the training process, the gradient normalization (GradNorm) based training strategy is leveraged to ensure adaptive balancing of loss decreasing. The extensive experimental results demonstrate that the proposed method significantly enhances transmission performance while maintaining relatively low computational complexity. The source codes are available at: https://github.com/llsurreal919/ADMIT.