Improved SSD network for fast concealed object detection and recognition in passive terahertz security images

Area of Science:

• * Computer Vision
• * Machine Learning
• * Security Technology

Background:

• * Global anti-terrorist measures necessitate effective security checks in public spaces.
• * Deep learning shows promise for detecting concealed objects in passive terahertz images.
• * Existing methods lack superior accuracy and real-time performance for labeling concealed threats.

Purpose of the Study:

• * To develop a novel method for accurate and real-time detection of concealed objects in terahertz images.
• * To enhance the performance of object detection algorithms for passive terahertz imaging.

Main Methods:

• * A deep residual network (ResNet-50) replaced the backbone of the Single Shot MultiBox Detector (SSD).
• * A feature fusion technique was employed to address missed or repeated detection of small targets.
• * A hybrid attention mechanism and Focal Loss function were integrated to improve detail acquisition, location accuracy, and model robustness.

Main Results:

• * The proposed SSD-ResNet-50 method achieved a mean average precision of 99.92%, an F1 score of 0.98, and a prediction speed of 17 FPS.
• * Accuracy improved from 95.04% to 99.92% compared to the original SSD algorithm.
• * The method demonstrated superior performance in accuracy and speed compared to Faster RCNN, YOLO, and RetinaNet.

Conclusions:

• * The novel SSD-ResNet-50 approach offers a significant advancement in real-time concealed object detection using terahertz imaging.
• * This method provides a robust and accurate solution for security screening in public areas.
• * The findings offer a valuable technical reference for developing deep learning-based terahertz smart security systems.