One-stage Network For Prohibited Items Detection Within X-ray Images

X-ray images prohibited items detection refers to the use of computer vision technology to process X-ray security images to detect whether the images contains prohibited items.This method can assist security inspectors in checking over-checked baggage,reduce the work pressure of security inspectors,and accelerate the speed of security inspection.Therefore,Xray images prohibited items detection has attracted more and more researchers’ attention.Early researchers used traditional feature extraction algorithms to detect prohibited items in X-ray images,due to the chaotic background and high complexity of the images,they could not achieve better detection results.To this end,a single-stage dual network object detection algorithm based on deep learning is proposed to detect prohibited items in X-ray images.This algorithm is based on the single-stage object detection network YOLOv3,and makes two major improvements:(1)Due to the penetrability of X-rays,X-ray images are different from natural images.The items in the X-ray image are mixed with each other.In addition,due to the false color of the X-ray images,the entire images is very chaotic,and it is difficult to distinguish the background of the image from the objects to be detected.To this end,a DarkNet-C composite backbone network is proposed,which consists of two DarkNet-53 basic backbone networks,which are used as assiatant backbone networks and lead backbone networks.The feature information of the image is spread backwards layer by layer through the cross-layer splicing between the two basic backbone networks,deepening the fusion between the shallow features and the deep features,so that the deep features still have the shallow feature information of the image,for example The shape and color information of the prohibited items in the image,so that the feature map obtained after going through the DarkNet-C backbone network still contains accurate image feature information.(2)In order to further improve the expression ability of features and improve the detection accuracy of small objects,the FAB feature Augment block is proposed.This block first splices the feature maps of the front and back layers of the network structure to improve the non-linear expression ability of features.Then the fused features are passed through the smoothing layer to eliminate the aliasing effect caused by feature fusion.By training the improved algorithm on the X-ray image dataset,it is proved that the detection accuracy of the algorithm is significantly improved compared with the existing methods,and it can complete the X-ray image detection task in real time.