Study on insulator recognition method based on simulated samples expansion

In the application of the deep learning to the unmanned aerial vehicle (UAV) autonomous inspection, a problem about the insufficiency of both quantity and quality for insulator images emerges. In the light of this situation, a sample expansion method based on a combination of statute and 3D modelling technology is proposed. Its feasibility is verified in a deep convolutional neural network by using five kinds of insulator simulated samples. The classification accuracy acquired by the proposed method is higher verified by the comparison of the experimental results, which proves its superiority to the traditional method containing no simulated samples. It is concluded that the simulated intensive samples of pure background have a significant effect on the accuracy of network classification. And when the ratio of real samples to simulated intensive samples goes to an appropriate value, the classification has the best accuracy.