Application of deep learning algorithms to terahertz images for detection of concealed objects

Sardar, S, Assi, S, Zolkifly, I.A, Jayabalan, M ORCID: 0000-0002-1599-965X, Alsaleem, M, Mohammed, A.H and Al-Jumeily OBE, D (2024) 'Application of deep learning algorithms to terahertz images for detection of concealed objects.' In: Bee Wah, Y, Al-Jumeily OBE, D and Berry, M.W, eds. Data science and emerging technologies. DaSET 2023. Lecture notes on data engineering and communications technologies (191). Springer, Singapore, pp. 279-289. ISBN 9789819702923

Abstract

Safety of the public at large venues is of utmost importance, and therefore, it is important to detect threats encountered due to concealed objects, especially on human bodies. Terahertz imaging has gained popularity over the last few years due to its ability to detect concealed objects inside fabric without harm to humans or invasion to their privacy. However, terahertz images suffer from poor resolution and low signal-to-noise ratio. Deep learning has shown high performance in classifying images lately, especially feedforward neural networks. Therefore, this study utilised deep learning for the detection of concealed objects on human bodies using the public active terahertz imaging (ATZ) dataset that contained images of 11 tiny objects. An end-to-end framework was applied and involved image enhancement using wavelet filtering, locating object using skip attention generative adversarial networks (SAGAN) and forward-forward mixed convolution with Gaussian affinity network (FMCGNet). The results showed that wavelet filtering did not affect the performance of the deep learning models. The SAGAN showed performance accuracy of 68.4% and that was related to the insufficient training of the model that required high computational power. However, the accuracy of classifying images with anomalies was high when FMCGNet was applied and was featured in the true positive rate of 91.9% though it was applied to imbalanced dataset. In summary, the proposed end-to-end framework successfully identified concealed objects on human bodies in an efficient way. Future work involves adopting this approach to different types of imaging datasets to further understand its strengths and limitations.

Item Type:	Book Chapter or Section
Divisions:	Bath School of Design
Date Deposited:	22 Aug 2025 15:29
Last Modified:	22 Aug 2025 15:29
URN:	https://researchspace.bathspa.ac.uk/id/eprint/17220

	Request a change to this item or report an issue
	Update item (repository staff only)