Research And Improvement Of A Dangerous Goods Detection Model Based On YOLOv3

Security has a close relationship with public safety,social stability and people's safety of life and property.Its timeliness and accuracy are of particular importance.In many public places such as railway station and airport,there is a so large flow of people and a variety of carry on luggage and personal effects that it has a high pressure in security.Up to now,the main method of security check at home and abroad is manual check assisted by X-ray security instrument.Since the security stuff identify dangerous goods by the eyes from X-ray images,this method is time and labor consuming and has a poor reliability.In the thesis,based on the characteristics and workflow of X-ray security instruments and considering the state of the arts in object detection,we connect dangerous goods detection with the object detection model based on deep learning to detect dangerous goods automatically with the deep learning network which can realize feature extraction,classification and positioning of the dangerous objects.Premised on the timeliness and giving priority to the accuracy of the detection,YOLOv3 is chosen as the basic model to study and improve.With the improvement in data preprocessing,network structure and strategy of choosing bounding boxes,we raise the Average Precision(AP)and mean Average Precision(mAP)of YOLOv3.The main research contents of the thesis are shown as follows:First,in view of the characteristics of the data set used in this thesis,we pre-process training data with the k-fold cross-validation method aiming to train the model adequately and suppress the over-fitting caused by insufficient quantity and uneven distribution of the training samples.Second,as vanishing gradient problem widely exists in the process of training deep learning models,it is a hot topic that how to suppress the gradient disappearing further,make full use of the shallow feature,enhance the feature reuse and reduce the training parameters in the improvement of algorithm and model.The dense connection structure of DenseNet is applied to darknet-53,the basic feature extraction network of YOLOv3.By changing some residual blocks in the original network to a dense block,the utilization of superficial features is enhanced,the number of network parameters is reduced,gradient vanishing is further suppressed and the detection accuracy is increased.Third,since object detection model based on the deep learning usually generates multiple regression bounding boxes,box selection algorithm affects the accuracy of the position detection,and then the precision of object detection.In the thesis,we improve the box selection algorithm of YOLOv3 by replacing NMS with soft-NMS,which reduces the risk of bounding boxes' false deletion and the mistaken delete of objects.By contrast experiments,it is found that compared with detection results of YOLOv3 on the image set from X-ray security instrument,the improved YOLOv3 can enhance both mAP and APs of some targets.It proves that the improved YOLOv3 has certain theoretical and practical value for real-time dangerous object detection,which contributes to reducing the pressure of security stuff,saving expenses and improving the efficiency of security check.