Deep Joint Source Channel Coding: A Novel Deep Learning Architecture with Hybrid Attention Modules

With the rapid advancement of deep learning techniques, source-channel coding and decoding methods are increasingly transitioning from traditional handcrafted designs to end-to-end learning frameworks. While existing deep learning approaches have demonstrated superior performance relative to conventional methods, they continue to face challenges related to performance instability in low signal-to-noise ratio (SNR) environments. To mitigate these challenges, this paper proposes a novel joint source-channel coding and decoding framework that incorporates transfer learning, spatial attention mechanisms, and channel attention mechanisms. Specifically, input image features are initially extracted using a transfer learning module, followed by the enhancement of salient information through spatial and channel attention layers. After simulating transmission over an additive white Gaussian noise (AWGN) channel, the image data is reconstructed using the same attention mechanisms and transfer learning module. Experimental results demonstrate that the proposed method outperforms existing approaches in low SNR scenarios, exhibiting robust performance and highlighting its potential applicability in complex channel environments.