双重对抗无监督域自适应绝缘子检测算法

绝缘子检测在输电线路智能巡检中具有重要的应用价值.基于深度学习的绝缘子检测是一类常用的方法.然而,在某些情况下仅能获取某一类型绝缘子数据,用其训练得到的模型直接应用到跨域绝缘子检测,性能会急剧下降.为此,提出一种双重对抗的无监督域自适应绝缘子检测算法.具体地,为缓解绝缘子图像背景复杂对检测性能带来的影响,设计了一种混淆判别机制.在该机制中,输入两种不同类型的绝缘子图像到两个不同的判别器中进行分类,再通过对抗训练将两种绝缘子进行交叉分类以学习到域不变特征.此外,通过最大最小化目标域的两个分类结果分别优化判别器和特征提取器,减轻不同类型绝缘子外观差异较大的问题.大量的实验证明了提出方法的有效性.

Unsupervised domain adaptation insulator detection algorithm based on dual adversarial

Insulator detection has important application value in the transmission line intelligent inspection, and insulator detection based on deep learning is a commonly used method. However, in some cases, only data of a certain type of insulator can be obtained, and if the model obtained by the training is directly applied to the detection of cross-domain insulators, its performance will decrease sharply. To solve this problem, a dual adversarial unsupervised domain adaptation insulator detection algorithm was proposed. Specifically, in order to reduce the impact of the complicated background of the insulator image on the detection performance, a confusion discrimination mechanism was designed, in which two different types of insulator images are input to two different discriminators for classification, and then the two insulators are cross-classified through adversarial training to learn domain-invariant features. In addition, the discriminator and the feature extractor were optimized respectively by the two classification results of the maximum and minimum target domains to alleviate the problem of large differences in the appearance of different types of insulators. A large number of experiments have proved the effectiveness of the proposed method.