Abstract
With the increasing demand for multimedia transmission over wireless networks, ensuring efficient and reliable image delivery is increasingly challenging due to channel impairments and limited bandwidth. Although semantic-aware deep joint source-channel coding (DeepJSCC) has emerged as a promising solution, most existing methods neglect the potential role of semantic correlation in rate control. To address this limitation, we propose a resilient deep coding neural network (RDCNN) that leverages the information bottleneck (IB) principle for rate optimization. The core of our approach is an improved rate-distortion (RD) guided image semantic communication system, which comprehensively considers the rate, pixel and semantic-level distortion, and semantic entropy. A rate-adaptive mechanism is constructed, which adjusts based on channel conditions, using Gumbel-Softmax sampling and a weighted asymmetric Gaussian distribution. Furthermore, the framework incorporates a reconfigurable neural network architecture that adapts its complexity to the signal-to-noise ratio (SNR), thereby achieving a balance between computational complexity and transmission accuracy. Simulations across datasets and channel conditions demonstrate that RDCNN consistently outperforms conventional methods and state-of-the-art baselines in terms of channel bandwidth ratio (CBR) and reconstruction quality. In particular, RDCNN achieves lower CBR while preserving high image fidelity, even under adverse channel conditions.