Hazardous object detection method based on improved YOLOv8 for terahertz security inspection equipment

The existing terahertz scanning equipment relies heavily on persons’ experience to identify image object. Because the resolution of the terahertz image is not ideal, the work intensity of the security personnel is very high, easy to get distracted or tired, so that long-term accuracy cannot be guaranteed. Therefore, it is of great necessity to intelligentize terahertz security inspection equipment to reduce manual labor intensity by deep learning technique. In this paper, we develop an automatic detection method of hazardous objects based on improved YOLOv8 for a terahertz security inspection equipment. The method realizes the automatic detection of dangerous objects by the improved YOLOv8 model. Specifically, the method incorporates Context Aggregation Networks into the YOLOv8 model to enhance its capability of feature extraction. To adapt to the low resolution of terahertz images, the neck network of YOLOv8 is designed as BiFPN. Additionally, the original C2f residual module is replaced with the C3 module to reduce model parameters and complexity, decreasing computational demands and increasing detection speed. Finally, EIoU is set as the target for model optimization. The experimental results show that the improved YOLOv8 model achieves a 95.2% mAP0.5 and 79.3% mAP0.5-0.95. The computational power requirement of the model is as low as 7 FLOPs and inference time is as fast as 1.4ms. With lower parameters and computations, the improved YOLOv8 model realizes improved detection accuracy and speed, outperforming current mainstream models including Sparse R-CNN, YOLOv5, and SSD, etc.