Research On Detection Method Of Fittings And Bolts Based On Faster R-CNN

There are a large number of fittings and massive bolts in transmission lines.Because they are operated outdoors all year round,it is easy to cause displacement,skew,damage,and missing bolts of the fittings,which threatens the normal operation of the power grid.The computer vision method was used to accurately and quickly detect the fittings and bolts in the aerial transmission line images.It is the premise for the condition and failure judgment of the fittings and bolts in the later stage.It is of great significance to guarantee the safe operation of the power grid.Aiming at the current situation that the existing Faster R-CNN object detection model based on deep learning has poor detection effect on hardware and bolts,this thesis presents the following work:In this paper,the RPN network in the Faster R-CNN model is improved to fit a variety of the fittings scale features,and the anchor generation method in the RPN network is modified.The detection accuracy is improved by 8% compared with the original Faster RCNN model,which makes the detection frame adapt to fittings object with different aspect ratios and scale ratios.At the same time,the Soft-NMS candidate box selection algorithm is used instead of the original candidate box selection algorithm.The problem that the detection frame cannot be accurately selected due to mutual occlusion of the fittings objects was solved,and the model detection accuracy was further improved by 1.41%.Through these two steps of improvement,the detection accuracy of Faster R-CNN model under Resnet-50 was finally improved to 79.76%.Aiming at the problem that the model cannot accurately predict the position of the detection frame due to the complicated position relationship between the fittings and the fittings and between the fittings and the background,this paper proposes a Faster R-CNN detection method that combines KL divergence and shape constraints.Based on the original Faster R-CNN detection model,the prediction branch of the objects bounding box distribution is added in this paper.The distance between the predicted distribution of the objects coordinates and the true value distribution is measured using KL divergence,and the model is optimized as a loss function.At the same time,the shape features of different the fittings objects are added as constraints to the loss function,which improves the generalization performance of the model and the accuracy of bounding box regression.To some extent,the problem of high detection score but large error in the position of the detection frame is solved,and the detection accuracy is further improved to 83.68%.Aiming at the problem that existing deep learning-based detection models cannot directly detect bolt objects in aerial images of transmission lines.This paper proposes a two-stage cascading detection framework.Firstly,the region of interest of the fittings is obtained from the Faster R-CNN typical fittings detection model combining KL divergence and shape constraints.Then use the FPN network,which is more sensitive to details,to further accurately identify and locate the bolts in the region of interest.The detection accuracy of bolts can reach 68.9%.While accurately retrieving the position of the bolt objects in the original transmission line inspection image in the original image,the high image resolution is maintained.The direct detection of bolt objects in aerial images of transmission lines is realized.