Improved Terahertz Image-Based Detection of Concealed Aviation Threats via DeepLabv3+ with DSConv and YOLOv5s

Abstract

In recent years, with the growth of civil aviation passenger volume, the importance of dangerous goods detection in ensuring airport and aviation safety has become increasingly prominent. To address the issue of low accuracy in existing dangerous goods detection methods, we propose an improved method for civil aviation dangerous goods detection. The method integrates an enhanced DeepLabv3+ model and a YOLOv5s model with an introduced Coordinate Attention (CA) mechanism. The DeepLabv3+ model is optimized by incorporating Depthwise Separable Convolution (DSConv) and Squeeze-and-Excitation (SE) attention mechanisms to enhance feature extraction capabilities. Meanwhile, the YOLOv5s model improves detection accuracy by incorporating the CA mechanism. We employ the HiXray, PIDray, and a self-collected terahertz dataset for training and validation. Using a methodology that involves semantic segmentation followed by object detection, experimental results demonstrate that the proposed method achieves an average precision (mAP@0.5) of 97.66% and a frame rate (FPS) of 27.23 f/s, outperforming comparison methods such as Faster R-CNN, EA-YOLOv8, and SSD. Additionally, an analysis of the application effectiveness of the proposed method shows detection accuracy rates of 97.8%, 96.6%, and 97.4% for knives, pistols, and lighters, respectively, with a CPU usage of 42.71% and a detection time of 17.89 ms, all of which are superior to the comparison methods. The above research results indicate that the developed hazardous material detection method is effective and practical. This method can provide a theoretical basis for research in the field of hazardous substance detection.