Smoother Soft-NMS for Overlapping Object Detection in X-Ray Images

摘要

As a contactless security technology, X-ray security inspection machine is widely used in the detection of dangerous object in all kinds of densely populated public places to ensure the safety. Unlike a natural image, various objects overlapping with each other can be observed in an X-ray image for its perspectivity. It brings us a challenge that the traditional NMS (Non-maximum suppression) algorithm will suppress the less significant objects. In this paper, we propose a Smoother Soft NMS based on the difference in aspect ratios and areas of different object bounding boxes to improve the accuracy of overlapping object detection. We also propose a special data augmentation method to simulate the generation of complex samples of overlapping objects. On our dataset, we boost the mean Average Precision of ResNet-101 FPN from 89.44% to 96.67% and Cascade R-CNN from 96.43% to 97.21%. Detector trained by Smoother Soft NMS has a significant improvement in overlapping cases.