Lightweight Real-Time Object Detection Based on Deep Learning.

Real-time object detection plays a significant role in the field of computer vision. Advanced object detection networks combine with the distribution characteristics in the image, while exposing detecting small targets as a bottleneck. In this paper, a novel network YOLO-light, can real-time detect and accurately detect objects in embedded system or portable devices. Firstly, in order to gaining better priori boxes, the clustering analysis is applied to pre-processing. Secondly, inspired from the multi-scale connection in the Feature Pyramid Networks (FPN) algorithm, YOLO-light enables multi-view and integrate features of various scales. With this design, YOLO-light can end-to-end training with low latency and higher average precision. The experiments testify that YOLO-light algorithm reveals satisfactory performance both in speed and accuracy.