A YOLOv5s-SE model for object detection in X-ray security images

Abstract

The target detection algorithm YOLOv5 is used in real-time object detection with its faster detection speed, good accuracy and flexibility of application. In this paper, a YOLOv5s-SE algorithm is proposed to solve the problem of complex background of X-ray security image and occlusion and overlap of objects. A custom channel attention module is added to the algorithm, which enables the network to focus on feature channels with a large amount of information and suppress unimportant feature channels. The position and quantity of the modules in the network are determined by the method of comparative experiments, and the GIoU loss of target box regression is replaced by CIoU loss to accelerate the convergence of the model. The object detection experiment on the X-ray security image dataset shows that compared with the YOLOv3 and Faster RCNN algorithms, the proposed algorithm has better detection results for partially occluded targets and the case where the distance between the targets is too close. Compared with YOLOv5s, the precision, recall, and mAP are increased by 1.2%, 2.8%, and 2.5% respectively, the loss convergence is also faster.