SNR-adaptive deep joint source-channel coding scheme for image semantic transmission with convolutional block attention module

With the development of deep learning(DL), joint source-channel coding(JSCC) solutions for end-to-end transmission have gained a lot of attention. Adaptive deep JSCC schemes support dynamically adjusting the rate according to different channel conditions during transmission, enhancing robustness in dynamic wireless environment. However, most of the existing adaptive JSCC schemes only consider different channel conditions, ignoring the different feature importance in the image processing and transmission. The uniform compression of different features in the image may result in the compromise of critical image details, particularly in low signal-to-noise ratio(SNR) scenarios. To address the above issues, in this paper, a dual attention mechanism is introduced and an SNR-adaptive deep JSCC mechanism with a convolutional block attention module(CBAM) is proposed, in which matrix operations are applied to features in spatial and channel dimensions respectively. The proposed solution concatenates the pooling feature with the SNR level and passes it sequentially through the channel attention network and spatial attention network to obtain the importance evaluation result. Experiments show that the proposed solution outperforms other baseline schemes in terms of peak SNR(PSNR) and structural similarity(SSIM), particularly in low SNR scenarios or when dealing with complex image content.