Multi-User Semantic Communications for Cooperative Object Identification

In this paper, a multi-user semantic communication system is studied to execute object-identification tasks, where correlated source data among different users is transmitted via a shared channel, and the introduced inter-user data-stream interference (IDI) deteriorates the identification performance severely. Traditional solutions adopt powerful channel codes for individual data protection, e.g., very low coding rate, to guarantee the identification performance, at the cost of sacrificing the real-time requirements. We propose to exploit the data correlation among users to perform cooperative identification. Specifically, by designing a convolutional neural network (CNN) based framework and constructing a combination of loss functions, a deep learning (DL) based multi-user semantic communication system for cooperative object identification, named DeepSC-COl, is proposed to fuse individual semantic features into a global feature through dynamically-tailored weights. In this way, multiple semantic features are jointly leveraged for identification without an extra increase of latency. Evaluation results show that the proposed DeepSC-COI outperforms the non-cooperative scheme with the performance gain of 86.9% at -3 dB, in terms of mean Average Precision (mAP).