Abstract: This paper presents a characteristic infrared thermal imaging combined with Softmax regression methods to identify insulator contamination levels. Through the infrared images of gray, image filtering, binarization, disk partition, semi-disk extraction pretreatment process, a single semi-insulator disk area is obtained. It designs with 6 characteristics of environment temperature, the average gray background image insulator, insulator disk region average gray, gray distribution of insulator disk, gray entropy variance and environment humidity that reflects contamination level based Softmax regression multiple classification model to identify insulator contamination levels. Probability threshold is introduced starting from the source of the problem to solve the resulting improper shooting invalid insulator infrared image on the contamination level of classification. Experimental results show that the selected feature sets and insulator contamination recognition model are efficient and feasible.

Key words: insulator, infrared thermal image, multi-class classification, Softmax regression, probability threshold

摘要： 提出了一种红外图像特征与Softmax回归相结合的方法识别绝缘子污秽等级。通过对红外图像的灰度化、图像滤波、二值化、盘面分割、半盘面提取等预处理过程，获取单个绝缘子半盘面区域。设计了以环境温度、绝缘子背景图像的平均灰度、绝缘子盘面区域的平均灰度、绝缘子盘面灰度分布的方差值、灰度熵和环境湿度共6个反映污秽等级的特征集的基于Softmax回归多值分类模型识别绝缘子污秽等级。引入概率阈值从问题源头出发，解决了拍摄时所产生的无效绝缘子红外图像对污秽等级分类的影响。实验结果表明所选取的特征集和绝缘子污秽识别模型高效且可行。

关键词: 绝缘子, 红外热像, 多元分类, Softmax回归, 概率阈值