Defect Detection of Anti-vibration Hammer Based on Improved Faster R-CNN

The damage or corrosion of the anti-vibration hammer will endanger the safe operation in the high-voltage transmission line. In this paper, the images of transmission line acquired during UAV patrol inspection are used as a research object. A deep learning method is proposed to improve the detection accuracy and defect identification of anti-vibration hammer in the influence of light variation, complex background, and small targets. Firstly, the original images are enhanced by using a Retinex algorithm to reduce the influence of light variation and shadow. Then the anti-vibration hammers are detected by using a deep learning framework called Faster Region-based Convolutional Neural Network (Faster R-CNN), in which the Feature Pyramid Networks (FPN) is used to extract and fuse the multi-scale feature of the image. A two-stage cascade Region Proposal Network (RPN) is designed to generate regions proposal. In the first stage, a standard RPN is used and the anchor is the proposal obtained by the sliding window. In the second stage RPN, the output proposal from the first RPN is as the new anchor position and a more accurate proposal can be obtained. Finally, the proposed boxes and the original feature map are sent to the subsequent network to complete the final defects classification and position regression of the anti-vibration hammer. Experimental results show that the proposed method improves the detection accuracy in 4 kinds of anti-vibration hammer defects, which has a good reference for the popularization of intelligent inspection of high-voltage transmission lines.