With the strengthening of global anti-terrorist measures, it is increasingly important to conduct security checks in public places to detect concealed objects carried by the human body. Research in recent years has shown that deep learning is helpful for detecting concealed objects in passive terahertz images. However, previous studies have failed to achieve superior accuracy and performance for real-time labeling. Our research aims to propose a novel method for accurate and real-time detection of concealed objects in terahertz images. To reach this goal we trained and tested a promising detector based on deep residual networks using human image data collected by passive terahertz devices. Specifically, we replaced the backbone network of the SSD (Single Shot MultiBox Detector) algorithm with a more representative residual network to reduce the difficulty of network training. Aiming at the problems of repeated detection and missed detection of small targets, a feature fusion-based terahertz image target detection algorithm was proposed. Furthermore, we introduced a hybrid attention mechanism in SSD to improve the algorithm's ability to acquire object details and location information. Finally, the Focal Loss function was introduced to improve the robustness of the model. Experimental results show that the accuracy of the SSD algorithm is improved from 95.04 to 99.92%. Compared with other current mainstream models, such as Faster RCNN, YOLO, and RetinaNet, the proposed method can maintain high detection accuracy at a faster speed. This proposed method based on SSD achieves a mean average precision of 99.92%, an F1 score of 0.98, and a prediction speed of 17 FPS on the validation subset. This proposed method based on SSD-ResNet-50 can provide a technical reference for the application and development of deep learning technology in terahertz smart security systems. In the future, it can be widely used in some public scenarios with real-time security inspection requirements.
© 2022. The Author(s).