Detection And Localization Of Concealed Objects In Millimeter Wave Images Using Deep Learning Neural Network And Image Inpainting

Because millimeter wave can penetrate human clothing to detect the concealed objects carried by human body,and it has no radiation harm to human body.Security inspection equipment using millimeter wave has been developed vigorously in recent years in China.The topic of this thesis is the detection and location of concealed objects in millimeter-wave human body images using the deep learning and image inpainting.There is no publicly available millimeter wave image data set around the world.The work of this thesis is based on the active millimeter wave image data set provided by the Shanghai Microsystem Institute of the Chinese Academy of Sciences.The main research work of this thesis is as follows.1)To obtain the millimeter-wave human completed images,we used the inpainting method based on DCGAN.DCGAN can learn the features of training set images and generate new images which are very close to training set images,so we can repair the original image with the new image.The exhaustive method divides the large area of the human body image into small areas,then digs out and repairs the small areas in turn.After the completion,a millimeter-wave human body image without prohibited items will be obtained.After preprocessing the data,constructing the DCGAN network,adjusting the parameters,the DCGAN network for millimeter wave human body images is successfully obtained using the Tensor Flow.Image completion is realized by using the DCGAN and exhaustive method.2)After completion,we need to compare the completed image with the original image to determine the presence of prohibited items,and the location of prohibited items when they exist.In the process of contrast,we use the SAR image change detection method and structural similarity index detection method.SAR image change detection is a point detection method,and the difference image is a two-dimensional matrix composed of likelihood ratio,which cannot accurately detect and locate prohibited items.Therefore,we adopt the regional judgment method,in which we find a region firstly,then set a threshold.If a certain proportion of points in the region are greater than this threshold,it is considered that this region contains prohibited items,so that prohibited items can be detected and located simultaneously.SSIM can be used to compare the similarity between two images or areas.The completed image is close to the original image,and does not contain prohibited items.Therefore,the two millimeter-wave human images before and after the completion can be compared in blocks using SSIM.If the SSIM is small,it indicates that the area may contain prohibited items.By analyzing the test results,it is found that the detection effect of SSIM detection method is better than that of SAR image change detection,which has lower false alarms when achieving the same detection power.Finally,we compare Faster RCNN's detection and location results with ours',and find that our method works better.