Semantic Transfer Between Different Tasks in the Semantic Communication System

At present, the semantic communication system considered the case that one of the intelligent tasks needs to be completed, while a single Internet of Things (IoT) device is usually required to complete multiple different tasks in reality. Moreover, it is difficult to obtain a large number of labels for some complex tasks to achieve high precision, such as object detection. Existing methods for few-shot detection mainly focus on transfer across domains, but there is the following problem in application: the parameters of the encoder cannot be shared at transmitter among different tasks, which leads to the storage pressure of the IoT devices required to transmit semantic information for multiple different tasks. To address this problem, we propose a novel transfer learning approach, Semantic Transfer Across Tasks, in which we leverage the semantic information to guide the training process of object detection with fewer labels and share the encoder architecture between classification and detection in the semantic communication system. Inspired by Grad-CAM, we select the semantic information which is important to detection, and we propose a semantic distance to improve the performance of few-shot detection. Experimental results show that our approach improved the mean average precision of few-shot detection in the semantic communication system and reduced the storage pressure of IoT devices.