Intelligent Contraband Recognition Algorithm And System Design For Millimeter Wave Security Equipment

Millimeter waves can penetrate clothes to obtain images of the human body and display contraband hidden under clothes.Therefore,the use of contactless millimeter wave security inspection equipment that can automatically identify contraband is increasing in domestic airports.However,due to the high noise and low resolution of millimeter-wave images,and the lack of imaging feature information of contraband,it is a very challenging task to accurately detect contraband in millimeter-wave images.At the same time,because the millimeter wave image will expose the privacy of the inspected person,the security inspector cannot see the image on the spot,and the security inspection equipment needs to recognize contrabands completely automatically,which requires high recognition accuracy.Currently,the recognition algorithms designed for visible light images are implemented directly on millimeter wave images.The imaging of millimeter-wave security inspection equipment has its own characteristics.It is the image of the human body in a fixed scene,and the imaging of contraband in different parts of the body has its own characteristics.Making good use of these characteristics can effectively improve the accuracy of contraband identification.Current recognition algorithms do not make good use of this information.Based on the imaging characteristics of millimeter-wave security inspection equipment,this thesis initially implements the millimeter-wave identification algorithm based on the classical two-stage object detection algorithm Faster Region-based Convolutional Neural Network（Faster R-CNN）to remove redundant candidates outside the human body through key point information,but the controllability of the candidate,flexibility cannot be achieved,so an novel two-stage contraband detection algorithm is proposed,called the Focal Detection（Focal Det）algorithm.In the first stage,the algorithm identifies the key point information of the human body,obtains the spatial structure information of the human body imaging,generates a detection focus covering the whole body,and generates a different candidate box for the contraband area on each focus.In the second stage,the algorithm determines the positive and negative sample label information of the candidate according to the area intersection of the candidate and the ground truth.By using various strategies such as smooth labels and smooth location,the information of each candidate is fully utilized to obtain more accurate contraband detection results.For millimeter-wave images,Faster R-CNN generates 9 anchor boxes with different sizes and shapes on each point of feature map.There are about 20,000 in total,and it is very time-consuming to filter these candidates.In contrast,the Focal Det algorithm proposed in this thesis generates candidate on the focal point,and only needs to generate85 regional candidates to cover the entire detection area.The candidate generation method of the algorithm simulates the experience of security inspectors during inspection,covering the whole body and checking key parts carefully.Region candidate boxes with different numbers,sizes,shapes,and center point offsets are generated on the focal points of different parts,and the generation method is more flexible.The generation method of the candidate makes full use of the image imaging characteristics and manual detection experience,so the algorithm model can learn and converge more efficiently.Compared with existing methods,the Focal Det algorithm has higher recognition performance,stronger generalization ability,and a good balance between precision and recall.The Average Precision（AP）value is 87.6%,which is higher than other comparative algorithm architectures,and the inference speed is 3 times that of the Faster R-CNN algorithm.With the outbreak of the epidemic,the demand for automated,digital,intelligent,and non-contact security inspections has become stronger.At present,millimeter-wave security inspection equipment is only an auxiliary role like a metal security inspection door,and it has not exerted its value.The problems to be solved in the design of the noncontact security inspection system include automatic scanning,privacy protection,mapping results to real pictures,and high recognition accuracy.The Focal Det algorithm proposed in this thesis can solve these problems well.First of all,the key point network part of the Focal Det algorithm can be independently used as a posture detection algorithm.It is judged that the posture of the inspected person is correct through the key points of the camera picture,and the millimeter wave device automatically scans.Secondly,input the millimeter wave image to the Focal Det algorithm,which can output the key points of the human body and the position coordinates of contraband.Through the key point information,the face,chest,crotch and other private parts can be accurately mosaicked,and the key points of the millimeter wave image and the key points of the real image of the camera can be registered,and the position of the contraband identified in the millimeter wave can be mapped to the real person.On the picture,you can see the hidden location of contraband in the real picture,which is like the perspective effect of penetrating clothes,and can also distinguish the misreporting caused by jewelry.Finally,through multi-dimensional information such as millimeter waves,cameras,and big data,it is judged whether the detected person has a threat,and then remote control or automatic pass conditions are set.The entire security inspection process is non-contact,and the test results are objective and not affected by subjective judgments.By mosaicking and mapping the recognition results to real images,the front-end will not expose millimeterwave human images,and privacy is fully protected.This thesis uses offline data to go through the process of the non-contact security inspection system and verifies the technical feasibility.The application of this system has important social significance.