Segmentation and Attention Network for Complicated X-ray Images

X-ray security automatic screening has attracted a broad attention for providing auxiliary support to eliminate potential threat in public. However, recent automatic detection algorithms may not perform quite well when a great number of dangerous X-ray images. segmentation consideration. articles are overlapping and compact in the In order to deal with it, this paper takes the scheme and attention network into Based on Mask R-CNN, we propose a two-stage deep CNN model called SAN to solve the performance degradation caused by the dense distribution of objects in X-ray images. The model contains a semantic segmentation network block to generate soft attention masks for each category and an attention network to combine each of the attention masks with the ROIs. In the experiment, the proposed SAN achieves 86% AP on our dataset, and brings an obvious improvement compared with recent detector algorithms on complicated X-ray dataset.