Prohibited items detection on X-ray images with multi-task learning

Inspection by X-ray has a wide range of applicability in airports, subways, railway stations, bus stations, and shipping ports. However, using X-ray images during security inspections is challenging because the X-ray image containing prohibited items (positive samples) in the real-world security inspection scenario is much smaller than those that do not contain prohibited items (negative samples). This results in low detection efficiency of the auto security inspection and many false detections. Considering this problem, this paper proposes detecting and recognizing prohibited items on X-ray images in real-world security inspections based on multi-task learning (MTL). This method uses an image classification network to filter out most X-ray images that do not contain prohibited items, thus reducing false detections and improving detection efficiency. MTL simultaneously trains the image classification network and the detection network of prohibited items. In addition, the backbone of the MTL network adopts a model sharing method that can reduce model parameters and improve the convergence speed of the model. Based on a dataset of security inspection X-ray images where the ratio of negative samples to positive samples is 100 (SIXray100), we performed data augmentation, estimated the target anchor of the training set, and used the concept of multi-scale training to complete the model training. Compared with the class-balanced hierarchical refinement method (CHR) that proposed a benchmark in detection, our method shows an improvement of 2% to 5%. Moreover, as the number of test images increases, our approach will obtain results that significantly exceed those from previous research. At the same time, this paper’s strategy for detecting prohibited items and its shared neural network framework have significant application value in engineering.