摘要
被动式太赫兹成像设备不断应用于人体安检,而在特定安检场合(例如机场)需要快速准确检测出人体安检图像中的小目标违禁品。由于被动式太赫兹图像低分辨率的限制,人体安检对小目标检测精度和速度的需求,传统检测方法和单纯的深度学习算法都难以满足。根据被动式太赫兹图像低分辨率的特性,本文在目标检测前对图像进行预处理。针对小目标检测,我们对SSD算法的网络结构和预选框进行了优化。在小目标检测方面,我们的方法可以提高检测精度到73%。而在一般目标检测方面(包含小目标)我们的方法可以达到76%mAP。
Passive terahertz imaging equipment is constantly used in human body security check. In the specific security situation such as airport, it is necessary to detect small object contraband quickly and accurately. Due to the low resolution of passive terahertz image, the traditional detection algorithm and simple deep learning algorithm cannot meet the needs of accuracy for small object detection in human body security check. In this paper, according to the low resolution of passive terahertz image, we preprocess the image before object detection. To improve detection accuracy for small objects, we optimize the architecture and default boxes of SSD model. For small object detection, our method, the optimized SSD with preprocessing, can improve accuracy to 73%. For common object detection, our method can achieve 76% mAp.
出处
《计算机科学与应用》
2020年第5期1064-1073,共10页
Computer Science and Application
