AC‐SDBSCAN: Toward concealed object detection of passive terahertz images

Abstract

The passive terahertz (THz) detection on objects concealed under clothing is quite challenging due to the heavy noise interference. In this paper, an active-contour-based self-adaptive DBSCAN (AC-SDBSCAN) detection algorithm is proposed. The core of AC-SDBSCAN algorithm lies in that the object contours are first extracted by AC method in a noise scenario and then the statistical features of the contours are used to motivate a SDBSCAN to complete clustering without initialization. Benefiting from the strong robustness of AC and DBSCAN to the noise, the concealed objects in the noisy THz images can be detected accurately. Extensive simulations are verified on four passive THz image datasets. The results indicate that the AC method in our solution can achieve over 87% accuracy for contour extraction of passive THz images, while the classical methods achieve less than 77%; in addition, the SDBSCAN can achieve over 90% clustering accuracy without manual initialization which is significantly superior to the conventional DBSCAN. Eventually, the proposed method completes the object detection of passive THz images with a maximum recall of 90.38% and a maximum precision of 94%.