Research On Insulator Defect Detection Based On Improved RetinaNet Algorithm

Most of the power supply accidents in the power grid are caused by insulator defects.Therefore,in order to ensure the continuous,stable and safe operation of power grid facilities,it is necessary to conduct regular inspections on the power grid.As the advantages of deep learning algorithms in the field of target detection have clearly surpassed traditional algorithms,it is of practical significance to apply deep learning algorithms to insulator defect detection tasks.Therefore,this paper studies two insulator defect detection schemes based on the deep learning algorithm:For the identification and detection of insulator defect images returned by UAV through ground server,this paper proposes an improved Retina Net algorithm.In order to further enhance the feature extraction capability of the Retina Net detection algorithm,this paper uses Rep VGG with stronger feature extraction capability instead of Res Net as the backbone network and combines the attention module to change the Io U to CIo U.Experiments show that the accuracy of the algorithm in insulator defect detection is improved by 3.5%compared with the original Retina Net algorithm,and the frame rate is nearly 15 frames per second.And the Raspberry Pi is used as the lower computer to simulate the UAV to transmit the signal to the ground server for detection and analysis.The results show that the frame rate of the ground detection terminal is basically the same as the frame rate of the direct use of the server for detection.For the direct detection of insulator defects through embedded devices,a lightweight detection algorithm is proposed in this paper.Since the overall structure of the Retina Net algorithm and the FCOS algorithm is the same,the main difference is that FCOS is an anchor-free target detection algorithm,and Retina Net is an anchor-based target detection algorithm.Benefiting from the speed advantage of the anchor-free model,this paper integrates the lightweight model Ghost Net on the basis of the FCOS detection algorithm,and removes the feature map of the FCOS algorithm for detecting large targets according to the actual situation of small insulator defects in the data set.The insulator defect detection accuracy of the improved lightweight model can reach 91.4%,and the average frame rate is nearly 40 frames per second,which is about 25 frames per second higher than the original RetinaNet algorithm.