Crack Detection of Electrical Equipment Based on Improved GoogLeNet

Abstract

Crack detection of electrical equipment is significant to maintain its normal operation. Many methods based on deep learning have been applied to detect cracks from the captured images, while most of the existing crack detection algorithms cannot detect the crack quickly and effectively, and rarely applied in electrical equipment with complex structures. In this paper, an improved GoogLeNet by combining DenseBlock and feature fusion layer is proposed. To reduce the amount of network training parameters, DenseBlock is utilized to replace the two branches with a large size convolution kernel in Inception model of the classical GoogLeNet. Moreover, to improve the detection accuracy of the network, a fusion layer integrating deep and shallow features is introduced in the improved GoogLeNet. To mitigate the issue of limited amount of training image data of electrical equipment, except for data augmentation, a transfer learning strategy is used to initialize the parameters of the improved GoogLeNet, where the initial parameters are obtained from the results of training public crack datasets. The experimental results show that the improved GoogLeNet can effectively detect the crack of electrical equipment, and the detection accuracy reaches 97.06%.