Research On Prohibited Items Detection Algorithm Of Dual-View X-ray Security Images

The current X-ray security checks are mainly by the naked eye of the security officers on the X-ray security machine scanned X-ray image identification,to determine whether there are prohibited items.Long work hours can lead to staff fatigue,resulting in the occurrence of missed checks,false checks and other situations.Many places have used dual-view X-ray screening machine,security inspectors can be more convenient and effective from two different view images to find contraband.With the wide application of computer vision technology in the security field,many deep learning-based detection frameworks have been used to assist security staff in quickly discovering contraband,but among these results,there is relatively little research on dual-view X-ray image detection algorithms;and there is a lack of publicly available dual-view datasets to support the research.Based on this,this thesis constructs a relevant dataset with dual-view X-ray security images,and combines deep learning to carefully analyze and study dual-view X-ray security image detection.The main work and innovations of this thesis include.(1)A dual-view X-ray security screening image dataset was constructed and refined.The required images were all taken using the latest dual-view X-ray security machines;each image was manually annotated by the volunteers.The dataset contains a total of 8742 images and corresponding correct annotations for six types of contraband(Knife,Pliers,Wrench,Scissors,Lighter,Powerbank).(2)Based on the above dataset,the performance difference of two types of target detection methods(one-stage,two-stage)in two different views of X-ray security images is explored.The YOLOv4,SSD and Faster R-CNN detection frameworks are trained and evaluated on the dataset constructed in this thesis to compare the accuracy of prohibited items according to different viewpoints,respectively.At the same time,the differences in detection accuracy between different viewpoints when the target is the same are analyzed in single-view view detection to provide a basis for comparison of data for subsequent studies of prohibited items detection algorithms for dual-view X-ray security images.(3)Based on deep learning,a dual-view X-ray security screening image detection algorithm is proposed.By combining the visual information of different viewpoint X-ray images in the dataset,the model improves the detection accuracy of prohibited items.And a fusion module is designed to filter the features in the horizontal view,while suppressing irrelevant noise and emphasizing the effective information.Finally,the processed horizontal view features are fused into the features of the vertical view,and the detection of contraband in the vertical view is improved by constructing the inherent spatial and logical information in both views.