Few-shot concealed object detection in sub-THz security images using improved pseudo-annotations.

Area of Science:

• Computer Vision
• Machine Learning
• Security Imaging

Background:

• Few-shot object detection is crucial for identifying concealed items in security imaging.
• Adapting machine learning for sub-terahertz (sub-THz) imaging presents challenges due to limited labeled data.

Purpose of the Study:

• To develop an effective few-shot object detection framework for concealed object identification in sub-terahertz images.
• To address the scarcity of labeled data in the sub-THz domain through innovative data augmentation.

Main Methods:

• Proposed a novel pseudo-annotation method to generate high-quality training samples from unlabeled sub-THz images.
• Employed fine-tuning based machine learning frameworks adapted for the sub-THz domain.
• Utilized multiple one-class detectors and a fine-grained classifier trained on thermal-infrared images to prevent overfitting.

Main Results:

• Successfully enhanced the detection of challenging concealed objects (e.g., 3D-printed guns, ceramic knives) with few-shot learning.
• Demonstrated improved model performance in real-world scenarios with scarce training examples.
• Validated the effectiveness of the pseudo-annotation technique for augmenting object detection capabilities.

Conclusions:

• The proposed pseudo-annotation method significantly improves few-shot object detection in sub-terahertz security imaging.
• This approach offers a viable solution for identifying dangerous concealed items when labeled data is limited.
• The study highlights the potential of adapted machine learning frameworks for advanced security applications.