A Multi-Task Semantic Segmentation Network for Threat Detection in X-Ray Security Images

Abstract

One of the key technologies of intelligent security inspection is threat detection. Existing threat detection methods for X-ray security images are mainly based on image classification and object detection. The prohibited items in X-ray security images have the problems of overlaps, blurred borders and different scales. To obtain more precise locations of prohibited items, we propose a multi-task semantic segmentation network to categorize prohibited items at the pixel level, so as to assist artificial security inspections more effectively.. For the limited dataset, we propose a simple and effective data augmentation method for X-ray security images. In the feature level, we propose a global multi-scale (GMC) module and combine the multi-level features in the decoder part. In the supervision level, we propose a multi-task loss using extra minimal bounding rectangles and edges as supervisions to strengthen the learning of the shape of prohibited items. Extensive experiments illustrate that our method achieves better performance on the prohibited items segmentation task. We achieve 71.06% mIoU accuracy and 83.28% pixel accuracy on JinNan X-ray Dataset.