Feature Maps Training for Substation Defect Detection
Xingyu Yan, Jie Na Zhou, Yanqing Ma, Ning Wang, Yuxing He, Hui Cao · 2021
To ensure the safe and reliable operation of substation, automatic detection of substation defects is the key of smart grid. Lightweight model is used in substation defect detection because of its small amount of parameters and calculation. However, the traditional lightweight model has the problem of low accuracy, which often leads to low recognition success rate. In this paper, a method of feature maps training is proposed. Because the feature map of complex network provides high-level knowledge, this training method takes the feature map of complex network as another training target. The actual defect image is taken as the experimental data set. The experimental results show that the accuracy of the network based on the feature maps training is higher than that of the traditional lightweight network, and it can effectively identify the substation defects.