Research On Security Inspection Based On Deep Learning

With the rapid development of deep learning in recent years,object detection algorithms based on deep learning have achieved considerable results in many fields such as vehicle detection,face recognition,medical imaging diagnosis,and security industry.The application of deep learning greatly shortens the detection time,improves the detection efficiency,and ensures the safety of personnel.However,at present,most of the security inspections are still traditional manual security inspections,and the emerging X-ray security inspection equipment on the market has a correct rate of identifying 80% of the main types of contraband,and there is a lack of samples of rare contraband(firearms,etc.).The detection accuracy and recall rate need to be improved.In view of the contraband objects in rail transit scenarios,combined with the corresponding X-ray data sets and prohibited species,YOLOv3 +(You only look once)is further obtained on the basis of YOLOv3.In terms of network structure,YOLOv3 +improves the original Darknet-53 network to obtain five sets of feature maps of different sizes and uses a multi-scale fusion method to form a feature pyramid for X-ray contraband detection;In terms of loss function,in order to avoid the problem of gradient disappearance,a TSE loss function is used and it can adjust the weight as the error changes,so that the model converges better.In addition,this article establishes and analyzes the contraband data set under the X-ray security inspection environment,and further improves the quality of the data set through image processing methods.Then for the problem of data imbalance,the method of TIP(Threat Image Project)is used to obtain a more balanced data set.In order to verify the superiority of the YOLOv3 + method,YOLOv3 + is compared with popular methods in terms of detection accuracy and speed.Experiments show that YOLOv3 + is improving accuracy,real-time detection is still possible.