基于迁移学习的小样本输电线路巡检图像处理方法

Defect Recognition Using Few-shot Learning and Transfer Learning for Transmission Line Inspection Images

电力巡检无人机提供大量的巡检图像,但由于电力设备故障稀少,其中只有少量的设备缺陷故障图像可供使用。为了提高设备缺陷故障识别精度并减少训练过拟合问题,首先介绍了基于迁移学习的小样本电力巡检图像处理方法,通过图像裁剪、翻转、旋转等数据增强技术对小样本图像进行扩充,同时采用生成对抗网络(GAN)来扩充基础样本;并使用迁移学习技术,将基于大规模图像数据的预训练深度卷积模型进行定制,调整该神经网络模型的输入层和最后两层参数,并对超参进行调优。实验结果表明,巡检设备故障(如导线断股和绝缘子串脱落等)准确匹配度近95%,证明了小样本学习和迁移学习在输电线路巡检图像处理中具有可行性。

A large number of images are provided with utilization of unmanned aerial vehicle(UAV)in grid inspection while few images of defected devices are available due to the nature of possibility of defects.To increase the defect recognition accuracy and alleviate the problem of overfitting,we introduce few-shot learning technology by applying data augmentation such as image crop,flip and rotation,etc.The generative adversarial network(GAN)is also used to generate samples for image data augmentation.We subsequently developed classification model for grid inspection defect recognition with transfer learning technology.We change neural network parameters on input layer and last two output layers,and leverage state-of-the-art pre-trained deep convolution neural networks models on large-scale data.Fine-tuned technology is applied as well.The experimental result shows the recognition accuracy of defects such as line broken and insulator broken is closed to 95%.It proves that few-shot learning and transfer learning can be effective approach for defect target recognition in transmission line inspection image processing.