In the power system, it is very important to identify and eliminate the hidden dangers of transmission lines to ensure the power system's security. Image recognition technology is an effective method to identify the risk of breaking out. According to the hidden breaking danger recognition problem, this study proposes a depth model by training the convolutional neural network algorithm. According to the anti breaking characteristics of risk image on the existing depth network structure are improved by increasing the ROI pool layer and modifying the loss function. A large number of training samples are used to get the robust model test when the measured image is first in generated candidate region, then the detection and identification for each candidate region are carried out, to detect potential risks to break out in a complex background. The experimental results show that this method can effectively identify the hidden danger of transmission lines.