Deep Joint Source-Channel Coding for Multi-Task Network

Multi-task learning (MTL) is an efficient way to improve the performance of related tasks by sharing knowledge. However, most existing MTL networks run on a single end and are not suitable for collaborative intelligence (CI) scenarios. In this work, we propose network with deep joint source-channel coding (JSCC) framework, which allows operating under CI We first feature fusion based (FFMNet) object detection semantic segmentation. Compared other networks, FFMNet gets higher fewer parameters. Then split into two parts, mobile device edge server respectively. The generated transmitted through wireless channel server. To reduce transmission overhead intermediate feature, JSCC designed. By combining together, whole model achieves 512 compression loss within 2% both tasks. At last, training noise, robust various conditions outperforms separate source scheme.