Item: Segmentation algorithms for automatic concealed object detection using Terahertz imaging
No Thumbnail Available
Date
2025-05-16
Researcher
Chandel, Sushmita
Supervisor
Bhatnagar, Gaurav
Journal Title
Journal ISSN
Volume Title
Publisher
Indian Institute of Technology, Jodhpur
Abstract
The safety and security of individuals and property is an indispensable issue for public venues, transportation hubs, and sensitive facilities. The increasing threat of concealed weapons and other valuable objects under human clothing, calls for sophisticated and reliable security screening technologies. Traditional methods like metal detectors and X-ray scanners face limitations in detecting non-metallic objects and can often be intrusive raising privacy concerns. Moreover, the growing diversity and ingenuity of concealed objects require advanced imaging systems that can address these challenges while ensuring public safety. This thesis presents novel approaches leveraging passive Terahertz (THz) imaging. It is a non-ionizing, privacy-preserving, covert, and contactless way of scanning humans for visualizing potential concealed objects. When applied to these images, image processing and computer vision algorithms give an automatic solution to the problem. They can provide a completely automatic solution to detect, localize, and recognize these objects in real time. This technology is in its infancy and has huge potential in the near future. This thesis attempts to solve the problem by taking a segmentation based approach to the problem. Unlike the visible spectrum which also captures an object’s color and inner texture, the object visualization in this spectrum captures only the shape, size, and morphology of the objects. Thus, pixel-level localization would be more appropriate. The goal of this thesis is to review and analyze various image processing and computer vision based algorithms suggested in the literature to perform automatic concealed object detection under human clothing for security check and entry control applications using passive Terahertz imaging, with a particular focus on investigating and analyzing image segmentation based approaches for the same. The most prominent methods in the literature are broadly classified into class-independent binary segmentation/foreground extraction methods and class-dependent object detection methods. The former methods give a pixel-level binary mask leading to the detection and localization of concealed objects in a class-independent manner. They thus can be used for all kinds of known and unknown object classes. On the other hand, the latter is more concerned with the recognition of specific classes apart from just detection and localization and mostly uses data-driven deep learning concepts to give optimal performance. The research undertaken is broadly divided into three parts: The first part (Chapter 2 and 3) aims to study all the class-independent binary segmentation methods for the effective detection and pixel-level localization of concealed objects on an extensive dataset, to propose and analyze novel segmentation based algorithms for the same. The method in Chapter 2 utilizes particular image properties of the dataset, principles of focus of attention, and superpixel segmentation to give a generic, image processing method, free of any kind of learning for the purpose. Additionally, robustness to noise is also inherently considered in the study. Thereafter, to make the approach robust to varying imaging conditions, the principles of machine learning are leveraged in Chapter 3 to suggest a more general solution to the problem. It is worth mentioning that the particular insights developed in Chapter 2 help us engineer robust features for this learning-based method in Chapter 3. The second part (Chapter 4) of the thesis takes an application-specific take on class-dependent object detection. For this, a novel blob-detection-based approach based on hierarchical segmentation has been proposed to give a region proposal technique, a starting point for many object detection algorithms. The proposed technique is a real-time technique, free of any kind of learning and utilizes specific image properties for the same. Finally, the third part (Chapter 5) of the thesis is more concerned with measuring the accuracy of superpixel segmentation used in the first part of the research. Apart from a general literature review, a novel, efficient, and mathematically rich information theoretic measure for the same has been proposed and analyzed. It can be used to evaluate the refinement accuracy of the superpixel segmentation and choose optimal parameters. In conclusion, a comprehensive study of segmentation and hierarchical segmentation based automatic, and accurate yet resource efficient methods for the application has been done. Additionally, all the novel segmentation algorithms proposed in this thesis are implemented, and empirically evaluated on a dataset in this thesis.
Description
Keywords
Citation
Chandel, Sushmita (2019).Segmentation algorithms for automatic concealed object detection using Terahertz imaging (Doctor's thesis). Indian Institute of Technology Jodhpur