Camouflaged people detection based on a semi-supervised search identification network
Yang Liu, Congqing Wang, Yunlan Zhou
Abstract
Automated detection of military people based on the images in different environments plays an important role in accurately completing military missions. With the equipment gradually moving towards intelligence, unmanned aerial vehicles (UAVs) will be widely used for integrated reconnaissance/attack in the future. The lightweight and compact design of the small UAV allows it to travel through dense forests and other environments to capture images with its convenient mobility. However, as the camouflage has been designed to blend in with surroundings, which greatly reduces the probability of the target being discovered. Moreover, the lack of training data on camouflaged people detection will inhibit the training of a deep model. To address these problems, a novel semi-supervised camouflaged military people detection network is proposed to automatically detect the target from the images. In this paper, the camouflaged object detection dataset (COD10K) is first supplemented according to our mission requirements, then the edge attention is utilized to enhance the boundaries based on search identification network. Further, a semi-supervised learning strategy is presented to take advantage of the unlabeled data which can alleviate insufficient data and improve the detection accuracy. Experiments demonstrate that the proposed semi-supervised search identification network (Semi-SINet) performs well in camouflaged people detection compared with other object detection methods.