Research On THz Imaging Human Security Check Algorithm Based On Convolutional Neural Network

THz waves can penetrate into non-metallic materials such as clothing and parcels,and there is no ionization hazard to the human body.Therefore,THz imaging has become an excellent human security check solution.Traditional algorithms for THz security images focus on denoising,enhancement,and dangerous objects segmentation.These algorithms have poor generalization performance when faced with active THz images with rich edges and gray information,and cannot automatically detect dangerous objects.Aiming at these problems,this work introduces an object detection algorithm based on convolutional neural network to identify and locate dangerous objects in THz human security images.The main contents of this thesis are organized as follows:(1)The training of the object detection algorithm requires a large amount of manually labeled data.With a small number of THz human security images,using open source software to label the dangerous objects,and building a THz human security image dataset of dangerous objects.Then through different image processing algorithms,it's possible to augment the data during online training,which provides sufficient data.(2)Introduce region-based object detection algorithm,Faster R-CNN,to automatically detect dangerous items in THz images.By statistically analyzing the characteristics of dangerous objects,the ratio of anchor in Faster R-CNN is improved,and more accurate test results are provided.(3)Aiming at the shortcomings of Faster R-CNN algorithm,a Top-Down structure is introduced to fuse the multi-scale features in the backbone and realize the algorithm's prediction on fine-grained features.At the same time,by making full use of the existing multi-scale features,a multi-scale feature region proposal network is proposed and the anchor configuration strategy is improved.Experiments show that the Top-Down structure can greatly improve the detection performance of the algorithm,and the detection accuracy rate can reach 92% when the Res Net50 backbone is used.(4)This thesis analyzes the reasons for the low recall rate of the test results and uses the characteristics of the multi-angle imaging of the THz security checkpoint,proposes to synthesize the multi-angle image detection results by the threshold judgment,which effectively reduces the false and missed detection for the final check results.The effect of the results enhances the error tolerance of single image detection.The dangerous objects detection algorithm in this thesis can accurately detect four types of dangerous objects such as knives,bottles,rifles and guns in active THz images,and has great potential in human security check application.