X-YOLO: A deep learning based toolset with multiple optimization strategies for contraband detection

Observing X-ray images manually is a common method for detecting contraband in packages. Long-term continuous observation is prone to visual fatigue, leading to problems of missed detection and false detection. Motivated by aiding operators in contraband detection in packages, we propose X-YOLO which is a deep learning based toolset with multiple optimization strategies for contraband detection to increase the detection precision. The path enhancement is designed to shorten the information path between the lower layer and the uppermost layer. We replace Leaky ReLU with Swish and Steps with SGDR to make training stable. Mixup, a dataset-independent method for data augmentation, is designed to increase the amount of training data for improving the generalization of model without expert knowledge. In order to solve the issue that Intersection over Union (IoU) can not deal with two non-overlapping objects, we apply Generalized Intersection over Union (GIoU) as bounding box losses. The experimental results show that X-YOLO achieves mAP up to 96.02% and recall up to 98.55%, surpassing Faster R-CNN, SSD, YOLOv1, YOLOv2, Tiny-YOLO, YOLOv3, YOLOv3-tiny, YOLOv3-spp and YOLOv3 with some of optimization strategies.