| Recently,detecting and classifying concealed objects carried by passengers or visitors in airports and stations,or other public places,is a challenging work to local security check departments.Detection and classification of the objects in the images obtained by millimeter-wave imaging is one of the main methods to the challenge.However,due to low resolution,low signal to noise ratio,and low contrast of the millimeter-wave images,the automatic detection and localization of concealed objects in the images still remain as a difficult problem to be solved.In this dissertation,we study the problem of detection and localization of the concealed objects in active millimeter(AMMW)images based on deep learning.The contributions of the dissertation are summarized as follows:1)For the single-object scenario in the images,a low-complexity method is proposed for detection of the concealed objects in AMMW images.To our knowledge,object detection was performed by the sliding window algorithm with small size windows to match the small objects when dealing with millimeter-wave images.However,small size windows contain less contextual information,which cause high false alarm rate.On the other hand,a huge computational cost would be led when sliding a small window over the whole image.The main idea of the proposed method is of two steps.In the first step,a big window is employed to slide from the top to the bottom of the image,and the region of interest(Ro I)is generated.In the second step,by sliding another window from the left to the right of the Ro I,the location of the object can be obtained.The new method reduces the false alarm rate as well as the missed detection rate of the small size objects.An advantage of the method is that the computing time can be reduced by 70%.2)For the multiple-object scenario in the images,an extension of the Faster RCNN method is proposed to reduce the false alarm rate.Since the Faster RCNN uses the small region proposal window to match the small size object when applied in AMMW images,misclassification of human body parts by the detector occurs.To overcome this drawback,by adding a rechecking module to the Faster RCNN,the results of the region proposal network can be combined with the contextual information to recheck the final results.In this way,the false alarm from human body parts are filtered effectively,which highly improves the detection performance.3)A new detection method by semantic segmentation is proposed to improve the performance of object localization.Since the direct detection method usually employs a bounding box as the label of an object,which contains a large amount of human body background noise,the localization performance is certainly reduced.To overcome this problem,a semantic segmentation network is proposed,which is mainly based on dilated convolutional neural network.Since the dilated convolutional neural network can efficiently enlarge the receptive field while keeping the resolution of the feature maps unchanged,it is possible to segment the small-size objects and suppress the interference from human body parts simultaneously.The experiment results show that the performance can be improved by 38% in terms of the accurate localization metric.4)A pixel level automatic labeling method is studied for two objects in AMMW images.In fact,for object detection,only size and location information of each object are included in the training dataset.Relabeling at pixel level is required if the detection is formulated by segmentation,which makes the labeling work heavier.Here,a pixel level segmentation method is proposed for knife and gun.The proposed method employs the OSTU algorithm for image binarization,CANNY operator for edge detection and Radon transform for line detection.The method combines the local information and global information for region growth and region segmentation.In this way,the object can be segmented from complex human body background.The relabeling workload can be reduced by about 45% using the proposed method. |
PDF Full Text Request