Research On Target Detection Method Of Passive Millimeter-Wave Image Based On Deep Learning

With the rapid increase in demand for security check and security monitoring technology,millimeter-wave/terahertz security devices have begun to be deployed in public safety places such as airports and subways.Due to the need to process massive millimeter-wave/terahertz security image data,the efficiency of manual security is no longer sufficient.Therefore,the use of artificial intelligence to detect and identify contraband in millimeter wave images has become a research hotspot.In this paper,the Faster R-CNN algorithm is applied to the detection and recognition of passive millimeter wave images.Based on the Faster R-CNN model,an improved algorithm based on Faster R-CNN target detection is designed.The main research work of this paper is as follows:(1)First introduced the basic theory of deep learning.The target detection algorithm based on deep learning is selected by Faster R-CNN algorithm to detect the target in passive millimeter wave image,and the algorithm is analyzed.For the low resolution of the passive millimeter wave image and the small target area,based on the Faster R-CNN algorithm,the RPN(Region Proposal Network)network with FPN network is used to replace the RPN network in the FasterR-CNN algorithm.Low-level feature fusion adds more detail and position information and semantic information to the predicted feature map.Finally,the RPN network algorithm with FPN network has an average accuracy of 83.7% in the simulated millimeter-wave radiation image test set,which is 2.6% higher than the Faster R-CNN algorithm.(2)In the passive millimeter wave image,there are problems that the positive and negative samples are extremely unbalanced and difficult to distinguish between difficult and easy samples.Although Faster R-CNN has a strategy to deal with sample imbalance,for cases where positive and negative samples exceed 1:1000,Faster RCNN forecast will be biased towards a larger number of samples,resulting in a decrease in detection accuracy.In view of this situation,this paper introduces the focal loss function(Focal Loss)on the loss function design without changing the original Faster R-CNN algorithm to deal with the imbalance of positive and negative samples.The loss function coordinately controls the loss weights of the positive and negative samples and the loss weights of the difficult and easy samples through the two parameters ? and ?,respectively,to ensure the loss balance of the positive and negative samples and the difficult/easy samples.The average improved accuracy of the final improved loss function algorithm in the simulated millimeter wave radiation image test set was 83.0%,which was 0.9% higher than the Faster R-CNN algorithm.(3)During the security inspection,the target contraband was not detected due to factors such as occlusion and deformation.In response to this problem,a stealth object imaging video(frame sequence)is formed by rotating a passive millimeter wave security detector 360 degrees to form a hidden object,and the video is detected and recognized in real time,and the detection results of each frame are combined to prevent missed detection and false detection.Finally,the improved algorithm proposed in this paper has an average accuracy of 84.3% in the passive millimeter wave radiation image test set,which is 2.2% higher than the Faster R-CNN algorithm;the video detection speed reaches 20 fps,which can be seen from the video detection effect chart.Effectively prevent false detection and missed inspection of security products.