Research On Prohibited Item Detection Method Based On Graph Matching

X-ray imaging technology has been widely used in security inspection.The application of this technology can effectively prevent criminals from carrying out terrorist attacks or illegal transactions.Compared with inspectors to determine whether there are prohibited items in X-ray images,automatic security check technology can effectively improve the efficiency of security inspection and save labour costs.Therefore,the detection of prohibited items in X-ray images based on computer vision and pattern recognition technology has an strong application demand.To solve the problem that objects present different scales,different rotation angles and occlusion in X-ray images,and the scarcity of prohibited items samples in X-ray images,our article designs graph matching algorithms for detecting the X-ray image of prohibited items according to the maximum similarity subgraph matching algorithm and the graph matching neural network algorithm.And the experimental verification and comparative analysis are carried out on the X-ray data sets.The main work of our thesis includes:(1)A prohibited item detection algorithm based on maximum similarity subgraph matching is proposed.This algorithm uses the structural attributes of the object which takes the structural features of the image as nodes,and the relative distance between nodes as edges to build the graph model.Then this algorithm gradually cuts down the nodes to carry out subgraph matching.Experimental results show that the proposed algorithm can detect prohibited objects at different scales,different rotation angles and occlusion without training samples and has a fast detection speed.(2)A prohibited item detection algorithm based on graph matching network is proposed.This algorithm takes the features extracted from the convolutional neural network as nodes to establish a full-connection graph model of images.Then our model classifies the target graph nodes by using the similarity information of nodes between graphs.Finally our model locates the target objects according to the relative positions of the classified graph nodes to realize multi scale,multi rotation and partial occlusion detection of prohibited items.Since the classification of the target graph node in our algorithm is independent of the object category of the input images,so it can effectively deal with the problem of scarce samples of prohibited items.Experimental results show that the proposed algorithm can significantly improve the performance of prohibited item detection under complex background and high occlusion rate.(3)An graph matching network detection algorithm based on deep EMD algorithm is proposed.On the basis of the graph matching network proposed in our thesis,this model introduces the graph similarity calculation method based on the deep EMD algorithm,so that the algorithm not only matches from the perspective of the node similarity of the graphs,but also establishes the global concept to improve the robustness of the graph nodes classification.Experimental results show that the proposed algorithm improves the detection performance by 1.5%compared with the original algorithm.The algorithm based on graph matching proposed in our thesis not only has scale invariance and rotation invariance,but also can adapt to the condition that the object is partial occlusion.In addition,the algorithm proposed in this thesis can reasonably solve the scarcity of the prohibited items samples in X-ray images.Experimental results show that the maximum similarity subgraph matching algorithm can effectively detect prohibited objects in X-ray datasets without training samples,while the graph matching neural network algorithms have better detection performance in the detection of X-ray images with complex background and high occlusion rate.