Semantic communication-based convolutional neural network for enhanced image classification

In AI-IoT environments, the traditional centralized cloud computing approach leads to high network transmission volumes and communication delays, negatively affecting intelligent task performance. This study addresses these issues by introducing an innovative semantic communication system model for intelligent tasks, leveraging deep learning techniques. The research focuses on image classification tasks constrained by bandwidth and delay in AI-IoT scenarios. The proposed model features a tailored semantic communication network architecture, where image feature maps are extracted on IoT devices. These semantic relations are then compressed based on the extracted feature maps to reduce power consumption on IoT devices and mitigate communication transmission pressures. Simulations and comparative analyses of various network performance metrics show that the proposed semantic communication system improves image classification accuracy by 90%% at low signal-to-noise ratios compared to traditional methods. With an 80%% compression ratio, the classification accuracy loss is minimal—within 2%%—when the signal-to-noise ratio exceeds 0. Additionally, at a signal-to-noise ratio of 20, the semantic compression transmission scheme enhances classification accuracy by 30%% compared to random compression. Moreover, the proposed system outperforms traditional approaches in execution time by approximately 80%%.