An Object Detection Network for Embedded System

Abstract

Object detection in images has a wide range of applications in various fields. However, many of the convolutional neural networks recently proposed have higher requirements on computing resources while achieving higher precision, which cannot guarantee good real-time performance on embedded platforms with limited resources. This paper proposed an object detection network suitable for embedded systems. The M-YOLO (Mobile-YOLO) model proposed in this paper combines depthwise separable convolution and residual blocks in feature extraction layers, which helps to reduce the amount of computation of the network. Multi-scale feature fusion is applied to the output layers to improve the accuracy. Experiments show that the M-YOLO model has 9.68M FLOPs (Floating Point Operations), which is about 22% of Tiny-YOLO model. The accuracy of the network reaches 56.61% on the PASCAL VOC dataset, and the speed in ARM is over 3 times faster than Tiny-YOLO model. The network is more suitable for embedded systems.