A Semantic Communication Approach for Multiscene Target Detection in Intelligent Vehicle Networks

As intelligent vehicle networks become integral to modern transportation systems, the need for accurate and efficient multi-scene target detection in complex urban environments is increasingly critical. Addressing the challenge of processing and transmitting extensive image data rapidly and accurately under dynamic road conditions, this paper proposes the Multi-scene Target Detection Semantic Communication (MTDSC) method. MTDSC employs convolutional neural networks and region proposal networks for effective target detection in images, and further utilizes spatial pyramid pooling and Long Short-Term Memory (LSTM) networks to assign relevant semantic labels in varied scenarios. These labels are transformed into a transmissible encoding format using a convolutional semantic encoder and a reinforcement learning strategy. The transmission phase leverages variational self-encoders techniques for channel encoding and decoding, ensuring reliable semantic information transfer in complex wireless environments. A crucial aspect of this research is semantic decoding and image reconstruction. To recover original semantic information accurately, a structured Support Vector Machine (SVM) loss function is utilized, designed to capture the structural relationships among labels and enhance the consistency between decoded and original semantic labels. Experimental results show that MTDSC effectively reconstructs semantic images across different road scenes, ensuring key information retention and high-quality image reconstruction, even in low bit rate and complex scenarios.