Abstract
Terahertz (THz) imaging provides a non-ionizing, safe alternative for security screening, yet it suffers from low signal-to-noise ratios and significant blurring. This paper proposes a detection framework combining mean-filter denoising with the Faster R-CNN architecture. By integrating localization and classification into a single deep learning pipeline, the system achieves an automated detection rate of 86.9% for concealed dangerous goods, such as knives, outperforming traditional computer vision methods.
Introduction
Security checkpoints traditionally rely on X-ray technology or manual inspections. While effective, X-rays pose radiation risks, and manual pat-downs are invasive and time-consuming. Terahertz waves, situated between microwaves and infrared light, can penetrate common materials like clothing without harming human tissue, making them ideal for non-contact security screening.
However, THz images are notoriously difficult to interpret. They often exhibit low resolution and high noise levels, requiring human operators to identify threats manually. This process is inefficient and prone to error. This research addresses these challenges by developing an automated detection system capable of identifying hidden objects within noisy THz data.
Methodology
The proposed system utilizes a two-stage approach to process THz images.
Image Preprocessing
To mitigate the inherent noise in THz captures, a mean filter is applied. This classic denoising technique smooths the image by averaging pixel values within a local neighborhood, effectively reducing graininess and preparing the image for feature extraction.
Faster R-CNN Architecture
The core of the detection system is the Faster R-CNN framework, utilizing a VGG16 backbone. Unlike older multi-stage detectors, Faster R-CNN unifies detection and recognition into a single network.
The architecture consists of four primary components:
- Convolutional Layers: These layers extract high-level feature maps from the denoised image.
- Region Proposal Network (RPN): A small network that slides over the feature maps to identify candidate regions (proposals) that likely contain objects.
- RoI Pooling: This layer extracts a fixed-size feature vector for each proposal, ensuring consistency regardless of the original region's dimensions.
- Classification and Regression: The final stage determines the object class (e.g., "knife") and refines the bounding box coordinates.
Experimental Results
The model was trained and validated on a custom dataset of 579 THz images (472 for training/testing and 107 for final proofing) featuring concealed knives.
The Faster R-CNN approach achieved a 86.9% recognition rate, successfully identifying 93 out of 107 target objects in the validation set. Comparative analysis showed that this method outperformed both traditional HOG-SVM (75.7%) and the single-stage YOLO detector (84.1%). While YOLO offered a slightly faster inference time (2.3s vs. 2.7s), the increased accuracy of Faster R-CNN is critical for security applications.
Conclusion
This study demonstrates that deep learning, specifically the Faster R-CNN architecture, can effectively automate the detection of dangerous goods in low-quality THz images. By combining traditional denoising with modern feature learning, the system provides a robust solution for non-invasive security screening. Future work will focus on expanding the dataset and extending the model to detect a broader range of threats, such as liquid explosives and narcotics.
Summary updated