Maximally Stable Extremal Regions for Concealed Object Detection in Passive Terahertz Imaging

In this work a method for concealed object detection using passive terahertz (THz) imaging has been proposed. Passive terahertz imaging is a promising technology for detecting hidden objects because it can penetrate fabric and identify various metallic and non-metallic items, such as weapons, explosives, drugs, and contraband, while providing high-resolution images. The detection methods studied till now in literature are either low-level intensity and spatial cues based class-independent image segmentation techniques or are based on class-dependent object detection techniques. The former techniques often break the concealed weapons into pieces due to the inherent noise in these images and are often too specific to the dataset and the operating frequencies. The latter techniques include both two-step detectors and one-step detectors. One-step detectors are deep learning based techniques requiring large amount of data, computation and class-information. Two step detectors employ either deep learning to give initial region proposals or use basic image segmentation techniques but in a hierarchical manner to give these regions. Sliding windows and selective search are the two very commonly used techniques using hierarchical approaches. This work proposes a novel technique to give region proposals, which are effective and address the particular problem of interest leading to fewer, faster yet accurate region proposals. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2025.