基于不确定性感知旋转目标检测的二次接线质检-《计算机技术与发展》

文章信息/Info

Title:: Uncertainty-aware Oriented Object Detection for Trustworthy Quality Inspection of Secondary Wiring

作者:: 毛泽勇1; 陈欣易1; 丁俊峰2; 陈蕾1; 2; 3*; 1. 南京邮电大学计算机学院,江苏南京 210023;2. 江苏省送变电有限公司,江苏南京 211106;3. 江苏省大数据安全与智能处理重点实验室,江苏南京 210023

Author(s):: MAO Ze-yong1; CHEN Xin-yi1; DING Jun-feng2; CHEN Lei1; 2; 3*; 1. School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;2. Jiangsu Power Transmission & Transformation Co. ,Ltd. ,Nanjing 211106,China;3. Jiangsu Key Laboratory of Big Data Security and Intelligent Processing,Nanjing 210023,China

Keywords:: oriented object detection; uncertainty-aware; Gaussian distribution modeling; secondary wiring quality inspection; YOLOv8

摘要:: 变电站二次接线质检旨在检查端子排上预设的接线编号与二次线缆线帽序号是否匹配。由于二次接线线帽尺寸小、分布密集且朝向各异的特点,传统的水平目标检测算法在该任务上表现不佳,而图像采集过程中的视觉畸变和噪声干扰则加剧了二次接线难例样本的误检。为此,提出一种不确定性感知的实时旋转目标检测算法(Uncertainty-aware Real-time Oriented Object Detection,UROD)并将其应用于变电站二次接线可信质检。具体地,基于 YOLOv8 算法引入角度回归分支以实现旋转目标检测功能,并对其边界框回归和分类分支分别进行高斯分布建模,UROD 能在输出目标检测结果的同时,伴随输出衡量检测结果的不确定性度量,而该不确定性度量又可应用于融合标识有序领域先验的接线成对匹配策略,从而实现二次接线难例样本的拒识。公开数据集 DOTAv1 与基于真实场景构建的二次接线数据集上的实验结果表明,相比于基线方法 YOLOv8,UROD 算法较大幅度提升了二次接线质检的精度;而相比于传统的旋转目标检测算法,UROD 算法则不仅提升了检测速度,而且能够基于其所感知的不确定性度量对难例样本进行拒识。

Abstract:: Secondary wiring quality inspection is designed to check whether the number on the terminal block matches the number of the wired cable cap. Due to the characteristics of small size,dense distribution and different orientations of wiring caps,horizontal object detection algorithm performs poorly on this task,while the noise interference during image acquisition exacerbate the misdetection of difficult case samples. To this end,an uncertainty -aware Real-time Oriented Object Detection ( UROD) algorithm is proposed and applied to the trusted quality inspection of substation secondary wiring. Specifically,based on the YOLOv8 algorithm that introduces an angular regression branch to realize the rotating object detection function, and models its regression and classification branches with Gaussian distributions. UROD can output the object detection results and its uncertainty metric. And the uncertainty metric can also be used in the wiring pairwise strategy. The experimental results on dataset DOTAv1 and the secondary wiring dataset show that compared to the baseline method YOLOv8, the UROD algorithm improves the accuracy of the secondary wiring quality inspection. Whereas compared to mainstream rotated object detection algorithms,UROD algorithm not only improves the detection speed,but also can reject difficult case samples based on the uncertainty.