Ghost Convolutional Neural Network-Based Lightweight Semantic Communications for Wireless Image Classification

Most convolutional neural network (CNN)-based lightweight semantic communication (SemCom) schemes mainly focus on reducing the number of regular convolutional modules to reduce the semantic encoder (SemEnc) complexity. However, this approach has limited ability to reduce the SemEnc complexity and weakens the representational capacity. To solve these issues, this letter proposes a ghost CNN (GCNN)-based lightweight SemCom scheme for wireless image classification. Specifically, we adopt the ghost convolutional (GC) module to extract semantic features, which reduces the SemEnc complexity and enhances the representational capacity. To prevent the gradient vanishing and improve the convergence speed, we utilize ghost bottleneck (G-bneck) blocks to stack GC modules. By cascading multiple G-bneck blocks, a lightweight SemEnc is constructed. Moreover, to enhance the robustness of the proposed GCNN against stochastic wireless channels, we design a spectral-spatial attention module that adaptively scales semantic features based on channel state information. Experimental results show that the proposed GCNN achieves the best classification accuracy and reduces the number of parameters and floating-point operations by factors of 2 and 8, respectively, compared with the state-of-the-art scheme.