A Split Edge Computing Doable Network for Object Detection base on Depthwise Separable Convolution

Due to the outstanding distributed data processing efficiency, edge computing has become a research hotspot in the field of object detection. Convolutional Neural Network (CNN) improves recognition performance of machine vision greatly, yet, which is difficult to deploy on edge computing device for its huge amount of data and calculation. Traditional deployment frameworks operate CNN completely on the cloud center or edge devices, while the cloud-only method leads to intolerable delay and bandwidth consumption, the edge-only causes the failure of edge devices when supporting massive computing tasks. In this paper, we propose an efficient edge computing framework and build a small target detection network: Split Edge Computing Doable Network (SECDN). In this framework, the feature extraction part is implemented on the edge device, and the parameters are compressed to avoid high calculation cost of edge devices. Raw data is preprocessed locally, and the results are sent to the cloud center for final processing. SECDN realizes the collaborative work of edge and cloud, and reduces pressure of edge or cloud. The experimental results show that the detection accuracy of SECDN has no obvious worse compared with the state of art network while requiring much lower data and computing effort.