A Deep Learning-based Contraband Detection Algorithm In Baggage Inspection Machine

As a non-contact security inspection equipment,dual-energy X-ray security inspection system has been widely used in public places such as rail transit,logistics,civil aviation and so on.However,the current X-ray security inspection work relies heavily on manual inspections,resulting in high labor costs but low detection efficiency.Lack of systematic training,the results of the tests from different security inspectors are different with each other.Further research has been carried out on the contraband detection at home and abroad recently,obtaining some results.However,overlapping objects in X-ray images still cannot be effectively detected.According to the practical needs of enterprises,a deep learning-based contraband detection algorithm is designed for luggage security inspection,achieving automatic detection of contraband.Firstly,the overall design of X-ray baggage security inspection of contraband is proposed and the contraband dataset is constructed.The needs of the contraband detection system is analyzed and the overall structure of the system is designed at first.Functional structure design of each module is put forward followed by the technical route of contraband detection algorithm design.Before analyzing the characteristics of the X-ray image,the X-ray contraband dataset is constructed,which lay the foundation for the design and improvement of the contraband detection algorithm.Aiming at the detection of contraband in the security inspection system,a contraband detection method based on Faster RCNN is proposed then.The down sampling structure,the feature fusion and the ROI Pooling method of the model are optimized for better detection.In order to reduce the difficult of network training,a reasonable set of Anchor is designed by Kmeans clustering algorithm.A comparative experiment with the one-stage contraband detector such as Retina Net and FCOS is designed to verify the effectiveness of Faster RCNN contraband detection algorithm.The improvement research of overlapping contraband detection algorithm based on Faster RCNN is carried out afterwards.In view of the lack of texture features in X-ray Images,a global context modeling module is established by combining the features of non-local areas for better feature extraction capability.According to the inaccurate positioning of the bounding box in the X-ray image,a reasonable box regression loss function is set up by combining the overlapping area,the center point distance,and the width-to-height ratio.In allusion to the problem of overlapping with other objects in the X-ray image,the data augmentation proposed increases the number of complex samples by generating some masks in the hazardous area,improving the performance of dangerous goods detection in overlapping scenes.For the overlapping characteristics of multiple dangerous goods in X-ray images,the improved nonmaximum suppression method takes the width-to-height ratio and area difference into account,which can better filter the boxes,reduce the suppression of overlapping objects and increase the recall rate of overlapping dangerous goods.The experimental results show that the improved contraband detection algorithm has an accuracy of 98.13% and 90.71% on the constructed dataset and the public dataset SIXray respectively,which greatly enhances the detection of overlapping contraband.Finally,the X-ray contraband detection algorithm module is deployed on the server through Tensor Flow Serving for the security inspection system to call,realizing the automatic detection of contraband.The deep learning contraband detection algorithm proposed in this paper has theoretical significance and application value for the detection of contraband in Xray images.It can replace the manual detection of contraband,which reduces the labor cost and improves the efficiency of security inspection.