Research On Insulator Recognition Algorithm Based On Deep Learning

Insulator is the key device of transmission line.In the transmission line,the number of insulators is large,widely distributed,long-term in the field environment,affected by high voltage and complex climate environment,prone to defects and cracks and other failures,once the failure will cause serious consequences and economic losses.Therefore,it is necessary to identify the insulator components in the aerial images efficiently and accurately,so as to provide the basis for the subsequent fault diagnosis and other related work.Accurate identification of insulator components is an important task of the combination of computer vision technology and electric power,and it is also the basic work to ensure the normal operation of transmission lines.Traditional insulator recognition algorithm is based on image processing technology,which needs to manually select features for recognition.Due to the complex background of insulator,easy to be blocked and other problems,the traditional method needs a lot of time and energy,which has been unable to meet the actual needs of power inspection insulator.With the development of deep learning technology,it is of great significance for scientific research and engineering to apply this technology to insulator recognition.The deep learning method can automatically identify the insulator target,so as to avoid the risk of high-altitude operation of inspectors,and also save a lot of manpower and material resources.This paper studies and analyzes a large number of insulator recognition methods based on deep learning,aiming at the problem that the semantic information of insulator in the recognition feature map is not rich enough,resulting in low recognition accuracy,an insulator recognition method based on feature integration and scaling network model is proposed.In the aspect of feature fusion,the feature map generated by backbone network is integrated from bottom to top;improving scaling module by 2×2 convolution kernel convolution,the integrated insulator feature map is input into the scaling module to generate four scale feature maps which are more conducive to insulator recognition;the anchor frame is generated and its proportion is adjusted.The insulator target recognition is completed through the normalized exponential function,and the frame regression is carried out to make the insulator location more accurate.The experimental results are compared with the traditional Faster RCNN,the improved Faster RCNN and the improved R-FCN,the average accuracy of this model is higher.In addition,due to the limitation of high computational complexity of the model proposed in this paper,the model will be further simplified by simplifying the model and optimizing algorithm in the follow-up work,so as to better integrate deep learning and power technology.