Dangerous goods detection based on transfer learning in X-ray images

Abstract

Computer vision technology is used to analyze X-ray images and detect dangerous goods in the process of logistics and express delivery. It is a security technology which can reduce labor strength and improve working efficiency. At present, there are many excellent detection models and methods in the field of object detection for visible light images, such as R-CNN, Fast R-CNN, Faster R-CNN, YOLO, SSD. These deep neural network-based detection methods achieved excellent performance on ImageNet. The training of object detection models on X-ray image datasets for dangerous goods detection is the focus of research in the field. Due to practical reasons, it is difficult to collect a comprehensive image dataset of dangerous goods (positive samples). In order to overcome this problem, this paper uses a multi-task transfer learning method on the basis of classification task and location search task on SSD network. The research in this paper focuses on adding additional convolutional layers in the SSD network to re-learn the knowledge of the model learned from the source domain. Experiments show that compared with the traditional method of fine-tuning, this method has better transfer learning ability on SSD network. This method was used to perform experiments in SSD300 on the image datasets screened from GDXray and achieved a mean average precision (mAP) of 0.915.