| 引用本文: | 赵振兵,李延旭,戚银城,孔英会,聂礼强.基于动态焦点损失函数和样本平衡方法的绝缘子缺陷检测方法[J].电力自动化设备,2020,40(0): |
| ZHAO Zhenbing,LI Yanxu,QI Yincheng,KONG Yinghui,NIE Liqiang.Insulator defect detection method based on dynamic focus loss function and sample balance method[J].Electric Power Automation Equipment,2020,40(0): |
|
| 摘要: |
| 在航拍输电线路图像的绝缘子缺陷检测任务中，针对不同类型缺陷之间存在的样本数量不平衡、困难样本低效学习等问题，提出一种动态焦点损失函数和一种基于二阶矩的样本平衡方法。首先在前向传播过程中根据困难样本、简单样本分布变化动态求解焦点损失函数的衰减因子，然后利用样本损失离散值定位出困难样本、简单样本的边界，从而获得困难样本集合，最后在反向传播过程中根据不同样本损失的二阶矩对学习样本的贡献率分布进行平衡。实验结果表明所提多类绝缘子缺陷检测方法能够有效地学习到不同样本的深度特征，性能较其他方法有显著的提升。 |
| 关键词: 多类绝缘子缺陷 样本平衡 损失函数 深度学习 目标检测 |
| DOI：10.16081/j.epae.202010008 |
| 分类号:TM726;TM216;TN911.73 |
| 基金项目:国家自然科学基金资助项目(61871182,61773160)；中央高校基本科研业务费专项资金资助项目(2018MS095,2020YJ006)；模式识别国家重点实验室开放课题基金(201900051)；北京市自然科学基金资助项目(4192055)；河北省自然科学基金资助项目(F2020502009) |
|
| Insulator defect detection method based on dynamic focus loss function and sample balance method |
|
ZHAO Zhenbing1, LI Yanxu1, QI Yincheng1, KONG Yinghui1, NIE Liqiang2
|
|
1.School of Electrical and Electronic Engineering, North China Electric Power University, Baoding 071003, China;2.School of Computer Science and Technology, Shandong University, Qingdao 266237, China
|
| Abstract: |
| In the task of insulator defect detection in aerial transmission lines, there are problems such as the sample numbers imbalance among different defect types and the inefficient learning of difficult samples, aiming at which, the dynamic focus loss function and the sample balance method based on second-order moments are proposed. Firstly, the attenuation factor of the focus loss function is dynamically solved according to the change of the difficult and simple samples distribution in the forward propagation. Then the discrete value of sample loss is used to locate the boundary of the difficult simple samples and obtain a difficult sample set. Finally, in the back propagation process, the contribution rate distribution of the learning samples is balanced according to the second-order moments of different sample losses. The experimental results show that the proposed multi-type insulator defect detection method can effectively learn the depth features of different samples, and its performance is significantly improved compared with other methods. |
| Key words: multi-type insulator defect samples balance loss function deep learning object detection |
