Robust Image Semantic Communication Based on Atrous Semantic Convolution and Adaptive Masking

Abstract

Typical wireless image transmission systems and deep joint source-channel coding frameworks often suffer from the “cliff effect” and feature homogeneity under adverse channel conditions. To address these limitations, we propose a robust image semantic communication system with semantic awareness and adaptive masking. Specifically, an atrous semantic convolution module is designed to capture the multi-scale context and precisely identify the region of interest within an image. Moreover, we present a mask adaptive module to dynamically adjust the mask ratios for different image patches by jointly considering the local semantic density and real-time signal-to-noise ratio (SNR), prioritizing allocating the limited bandwidth to the core semantic features. Furthermore, we integrate the ConvNeXt-based architecture with global response normalization to enhance the feature stability and prevent the representation collapse under high mask rates. Comprehensive experiments demonstrate that the proposed system significantly outperforms the wireless image transmission Transformer model, particularly in low-to-medium SNR environments. This system provides a robust and high-fidelity solution for semantic image transmission in complex and bandwidth-constrained scenarios.