Precise Localization of Concealed Objects in Millimeter-Wave Images via Semantic Segmentation

Existing concealed objects detection methods in active millimeter wave (AMMW) images are mainly based on bounding boxes. In this paper, we consider the problem of precise localization of concealed objects in AMMW images with the use of semantic segmentation networks. To improve the performance of the detection and localization of concealed objects, we propose a method with two steps. In the first step, we build a two-class semantic segmentation network to segment concealed objects in pixels from the images with the complex human body background, while in the second step, we use connected components extraction to detect and localize concealed objects in the segmented image. To improve the performance of the detection and localization of small objects, the network we built is composed of stacked dilated convolution blocks to enlarge the receptive field while keeping the resolution of associated feature maps unchanged. In addition, we give a rule for design of the associated dilation rates and the expand-contract dilation (ECD) assignment strategy for the pattern of the dilation rates. In the numerical experiments, we use the universal evaluation metrics, such as the AP (average precision) @ IoU (Intersection over Union)=0.5 and mIoU (mean value of IoU) to evaluate the performance of precise localization. The experiment results show that our method outperforms the existing ones for precise object localization in AMMW images, where the improvement of the AP@0.5 is about 38% and that of the mIoU is about 27%.