Hidden dangerous object detection for terahertz body security check images based on adaptive multi-scale decomposition convolution

Zijie Guo, Heng Wu, Shaojuan Luo, Genping Zhao, Chunhua He, Tao WangSignal 2025 · DOI

Abstract

• A deep learning hidden dangerous object detection method is proposed. • An AMDC module is developed to enhance feature extraction capabilities. • The proposed scheme has wide potential applications . • The proposed scheme can determine the category of hidden dangerous objects. • A model pruning approach is designed to optimize the proposed network. Recently, detecting hidden dangerous objects with the terahertz technique has attracted extensive attention. Many convolutional neural network-based object detection methods can achieve excellent results in common object detection. However, the existing object detection methods generally have low detection accuracy and large model parameter issues for hidden dangerous objects in terahertz body security check images due to the blurring and poor quality of terahertz images and ignoring the global context information. To address these issues, we propose an enhanced You Only Look Once network (YOLO-AMDC), which is integrated with an adaptive multi-scale large-kernel decomposition convolution (AMDC) module. Specifically, we design an AMDC module to enhance the feature expression ability of the YOLO framework. Moreover, we develop the Bi-Level Routing Attention (BRA) mechanism and a simple parameter-free attention module (SimAM) to integrate and utilize contextual information to improve the performance of dangerous object detection. Additionally, we adopt a model pruning approach to reduce the number of model parameters. The experimental results show that YOLO-AMDC outperforms other state-of-the-art methods. Compared with YOLOv8s, YOLO-AMDC reduces the parameters by 3.9 M and improves mAP@50 by 5 %. The detection performance is still competitive when the number of parameters is significantly reduced by model pruning.
BibTeX
@inproceedings{guo2025hidden,
  author = {Zijie Guo and Heng Wu and Shaojuan Luo and Genping Zhao and Chunhua He and Tao Wang},
  title = {{Hidden dangerous object detection for terahertz body security check images based on adaptive multi-scale decomposition convolution}},
  booktitle = {Signal},
  year = {2025},
  doi = {10.1016/j.image.2025.117323},
}

Loading related content...

Figure