难例挖掘在太赫兹成像目标检测中的应用-《计算机技术与发展》

文章信息/Info

Title:: Application of Hard Example Mining to Terahertz Imaging Object Detection

作者:: 薛飞¹; 2 ; 梁栋¹; 2 ; 喻洋³; 1. 南京航空航天大学计算机科学与技术学院模式分析与机器智能工业和信息化部重点实验室,江苏南京 211106;
2. 软件新技术与产业化协同创新中心,江苏南京 210093;
3. 中国工程物理研究院电子工程研究所,四川绵阳 621900

Author(s):: XUE Fei¹; 2 ; LIANG Dong¹; 2 ; YU Yang³; 1. MIIT Key Laboratory of Pattern Analysis and Machine Intelligence,School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;
2. Collaborative Innovation Center of Novel Software Technology and Industrialization,Nanjing 210093,China; 3. Institute of Electronic Engineering,China Academy of Engineering Physics,Mianyang 621900,China

Keywords:: terahertz image; THz Imaging; object detection; hard example mining; sample imbalance; RetinaNet

摘要:: 太赫兹成像中的隐蔽物体检测是公共安全和反恐的迫切需要。由于太赫兹成像质量差,在太赫兹图像上的目标检测比在计算机视觉领域常用的公共目标检测数据集上要困难得多。文中收集了一个多目标的主动太赫兹成像数据集。针对样本不平衡问题,对比了 RetinaNet 使用交叉熵和 Focal Loss 作为损失函数时的检测性能。针对那些检测效果较差的目标,利用难例挖掘技术来增强训练模型。由于传统的难例挖掘技术是在二阶段目标检测器基础上设计的,无法直接应用在一阶段检测器上,文章以 RetinaNet 为基础设计了一种以图像为单位的难例挖掘方案。实验也验证了 YOLOv3、YOLOv4、FRCN-OHEM 和基础的 RetinaNet 在该数据集上的性能。实验结果表明,Focal Loss 的使用提高了平均检测精度,难例挖掘技术的应用也提高了检测器对小目标等难例的检测率。

Abstract:: Hidden object detection in terahertz imaging is an urgent need for public safety and counter-terrorism. Due? ? to the poor quality of terahertz imaging,object detection on terahertz images is much more difficult than that? on public object detection datasets commonly used in computer vision. A multi - object active terahertz imaging data set is collected. The detection performance of RetinaNet when using cross entropy and Focal Loss as a loss function is compared for the sample imbalance problem. For those samples with poor detection performance, the training model is augmented using the hard example mining technique. Since the traditional hard example mining technique is designed on two-stage object detector and cannot be directly applied to a one-stage detector,a hard example mining scheme is designed based on RetinaNet with image as unit. The experiments also verify the performance of YOLOv3,YOLOv4,FRCN-OHEM and the base RetinaNet on this data set. The experimental results show that the use of Focal Loss improves the average detection accuracy, and the application of hard example mining technique also improves the detection accuracy of the detector for hard examples such as small objects.