Automatic Annotation Approach for Prohibited Item in X-ray Image based on PANet

In this paper, we propose an approach based on Polygon-Attention Network (PANet) for automatic annotation of prohibited item instances in X-ray images, aiming at accelerating annotation process for new datasets. X-ray image is fully special, mainly because it has a large amount of overlapping phenomenon, resulting in blurred boundaries of prohibited items. To solve this problem, we add an adaptive multi-level attention module to a generic encoder-decoder annotation architecture, which enables the model to adaptively fuse the required middle layer features in real-time according to characteristics of object boundary during each inference. To evaluate the proposed approach, we present a high-quality X-ray segmentation dataset named Prohibited Item X-ray (PIXray). Experimental results demonstrate that, through our approach, the annotation process speeds up by a factor of 2.4 in all classes in PIXray, and achieves an accord of 92.3% with ground-truth in IoU.