Enhanced visualisation of concealed target objects by infrared thermography and machine learning

• Infrared thermography and machine learning enabled accurate visualisation of concealed target objects. • Chan-Vese active contour algorithm and Fuzzy-c clustering produced accurate image segmentation. • Image fusion with visible spectrum image enabled enhanced interpretability of results. Electromagnetic waves such as millimetre-wave, terahertz, and infrared have been extensively employed in security scanning applications due to their non-invasive nature and strong detection capabilities. Millimetre-wave and terahertz imaging systems benefit from longer wavelengths, allowing for deeper penetration through clothing. This has enabled the development of commercial stand-off and walk-through detectors for the identification of concealed items beneath garments. In contrast, infrared radiation, while less penetrative and yielding weaker signals on clothing surfaces, offers significantly higher imaging resolution due to its shorter wavelength. Additionally, infrared cameras are generally more cost-effective, widely available, and well-suited for high-throughput applications over longer distances. In this work, a machine learning-based methodology was applied to thermal infrared images to improve the detection and visualisation of objects concealed under layered clothing. Principal Component Analysis was employed to identify pixels with marked thermal contrast between the subject and the background. This was followed by image segmentation using the Chan-Vese active contour algorithm and region-of-interest clustering using Fuzzy-c means. Finally, image fusion with the corresponding visible spectrum image was performed to enhance the interpretability of the results. Compared to K-means clustering, the proposed Fuzzy-c-based approach demonstrated improved performance in localising concealed objects, such as eliminating low heat signal due to the palm’s residual heat, underscoring the potential of machine learning-enhanced infrared imaging for practical security applications.

Sequential machine learning pipeline applied to thermal infrared images: PCA for thermal contrast identification, Chan-Vese active contour for segmentation, Fuzzy-c means for region-of-interest clustering, and image fusion with visible spectrum data for enhanced visualization.

Background & Context

Security screening for concealed threats has long relied on electromagnetic wave-based technologies like millimetre-wave and terahertz systems, which penetrate clothing effectively and have enabled commercial walk-through detectors. However, these longer-wavelength systems suffer from fundamental resolution limitations due to their physical properties.

Infrared radiation, while less penetrative and yielding weaker signals on clothing surfaces, offers significantly higher imaging resolution due to its shorter wavelength, and infrared cameras are generally more cost-effective, widely available, and well-suited for high-throughput applications over longer distances. This creates a compelling value proposition for security applications, but only if the weak thermal signatures of concealed objects can be reliably extracted and visualized.

This 2026 study by Khor and colleagues advances their earlier work on infrared-based concealed object detection by introducing a sophisticated machine learning pipeline specifically designed to enhance visualization under the challenging conditions of layered clothing. The research addresses a critical gap: transforming subtle thermal variations into actionable security intelligence without requiring the expensive, specialized equipment used in competing technologies.

Research Design & Methodology

The researchers applied a machine learning-based methodology to thermal infrared images to improve the detection and visualisation of objects concealed under layered clothing, employing Principal Component Analysis to identify pixels with marked thermal contrast between the subject and the background, followed by image segmentation using the Chan-Vese active contour algorithm and region-of-interest clustering using Fuzzy-c means.

The methodology employed a sequential processing pipeline designed to progressively isolate and enhance concealed object signatures:

• Principal Component Analysis (PCA) for initial feature extraction and identification of thermally distinct regions
• Chan-Vese active contour segmentation to delineate object boundaries without edge-based assumptions
• Fuzzy-c means clustering for region-of-interest classification with soft membership assignments
• Image fusion combining thermal and visible spectrum images for interpretability

The proposed approach was benchmarked against K-means clustering to evaluate relative performance in localising concealed objects. The study focused specifically on the challenge of detecting objects hidden beneath multiple layers of clothing, where thermal signals are particularly attenuated.

Results & Key Findings

Compared to K-means clustering, the proposed Fuzzy-c-based approach demonstrated improved performance in localising concealed objects, such as eliminating low heat signal due to the palm's residual heat. This represents a critical practical advance, as false positives from body heat signatures have been a persistent challenge in thermal-based detection systems.

Key technical achievements included:

• Superior segmentation accuracy — The Chan-Vese and Fuzzy-c combination outperformed traditional K-means clustering
• Artifact suppression — Successfully eliminated thermal noise from residual palm heat and other confounding sources
• Enhanced interpretability — Image fusion with visible spectrum data provided contextually meaningful output
• Layered clothing penetration — Maintained detection capability even under multiple fabric layers

The Chan-Vese active contour algorithm and Fuzzy-c clustering produced accurate image segmentation, and image fusion with visible spectrum image enabled enhanced interpretability of results. The integration of multiple algorithmic approaches created a robust pipeline capable of handling the complex thermal patterns generated by human subjects in realistic scenarios.

Discussion & Significance

This work represents an important evolutionary step in infrared-based security scanning, demonstrating that sophisticated image processing can partially compensate for infrared's inherent penetration limitations. The choice of Fuzzy-c means over K-means clustering is theoretically sound — fuzzy clustering's soft membership assignments are better suited to the gradual thermal gradients present in concealed object scenarios, rather than the hard boundaries assumed by K-means.

The successful elimination of palm heat artifacts is particularly significant for practical deployment. In real-world security contexts, subjects frequently touch clothing or carry items, creating transient thermal patterns that can trigger false alarms. The demonstrated ability to distinguish between residual surface heat and genuine concealed object signatures suggests the methodology could achieve acceptable false positive rates in operational settings.

Compared to the authors' earlier 2024 Scientific Reports study, which achieved area-under-curve values of 0.8923 to 0.9669 using CNN classification with various preprocessing approaches, this 2026 work focuses on visualization enhancement rather than binary classification. The complementary nature of these approaches — CNN-based detection for automated screening and enhanced visualization for operator interpretation — suggests potential for integrated multi-stage systems combining both methodologies.

Limitations & Considerations

While the study demonstrates clear advantages over K-means clustering, the research does not provide quantitative performance metrics such as detection rates, false positive rates, or processing time requirements. The effectiveness of the approach likely varies with fabric type, object thermal properties, environmental temperature, and time since concealment — variables that require systematic exploration.

The reliance on thermal contrast means the methodology may struggle with thermally neutral objects (those at equilibrium with body temperature) or in scenarios where subjects have recently been in cold environments. Additionally, the computational overhead of the sequential PCA → Chan-Vese → Fuzzy-c pipeline may present challenges for real-time walk-through screening applications, though this is not explicitly addressed.

Practical Applications

The findings underscore the potential of machine learning-enhanced infrared imaging for practical security applications, particularly in contexts where cost, deployment flexibility, and resolution are priorities. Airport security, government building checkpoints, event venues, and public transportation hubs could benefit from this technology as a complement to or replacement for more expensive millimetre-wave systems.

For security practitioners, this research suggests that infrared screening systems should incorporate sophisticated image processing pipelines rather than relying on raw thermal imagery. The visualization enhancement enables human operators to make more confident decisions, while the artifact suppression capabilities reduce operator fatigue from false alarms. The technology's non-invasive, privacy-preserving nature (thermal images do not reveal facial features clearly) may also facilitate public acceptance in civilian security contexts.