Brisk-Yolo: A Lightweight Object Detection Algorithm for Edge Devices

Abstract

YOLOv5 is a fast and accurate object detection algorithm which has been widely adopted in many industrial scenarios. However, it still has too much parameters and high computational costs for some edge and low-power devices. Thus, it is necessary to explore a more efficient object detection algorithm for cheap deployment. In this paper, we realize a more lightweight and efficient object detection algorithm based YOLOv5s, called Brisk-Yolo. Firstly, we construct a lightweight backbone based MobileNet V2, by introducing a sand glass block to replace inverted residual block at deep layers, which make the backbone more lightly with negligible loss of accuracy. Meanwhile, we use non-local block to replace SPPF block to help the model capture a wide range of spatial relationships with less parameters. Secondly, we propose an efficient and lightweight multi-scale feature fusion method "Cross-FPN", which combines FPN and an improved cross resolution weighting unit to promote the communication between multi-scale features. It gains accuracy improvement than original FPN and considerably reduces the number of parameters than PAFPN. Finally, a quality focal loss is introduced to boost accuracy further. Brisk-Yolo has only 3.68M parameters and 9.3 GFLOPS, which reduces about 49% parameters and 44% computation complexity than original YOLOv5s. Meanwhile, it achieves 34.1 mAP and 54.5 AP50 on COCO va12017 dataset, with a competitive trade-off between accuracy and model size.