Semi-Supervised Soft Label Assistance Network for Activate Millimeter Wave Concealed Object Detection
Yiran Shi, Zhenhong Chen, Dan Niu, Xin Wen, Zhenhang Pan, Mei Zhang · 2024
Activate millimeter wave (AMMW) scanners are widely used for the detection of concealed items. The mainstream AMMW detection methods are generally based on large labeled data. However, in the actual scenarios, it’s hard to obtain sufficient labeled images for training models. To overcome the problems above, this paper proposes a semi-supervised objection detection (SSOD) network with soft label assistance, which can achieve high accuracy in the case of only having a small number of labeled images. Firstly, a multi-scale attention method is designed to enhance the features of the region of interest and improve the detection accuracy of small targets. Secondly, a soft label-assisted strategy under the SSOD network is designed to capture key information in unlabeled data. Different from common pseudo labels, the soft label discards regular bounding boxes to mine foreground item information from the specific pixel level. It can not only extract crucial features outside the boxes but also eliminate redundant features inside. Finally, validation experiments on real AMMW datasets demonstrate the effectiveness of the proposed method.