Research On Fittings And Its Defect Detection Method Based On Automatic Framework Searching FPN

Transmission line inspection is the guarantee for the safe transmission of electric energy.Intelligence inspection technology based on computer vision is the development demand for transmission line inspection.As the most versatile and versatile functional metal component the on transmission line,the fittings are one of the most vulnerable components.The fittings and their defect targets are the main inspection targets in the transmission line inspection task.How to quickly and accurately locate the fitting targets and detect whether the fitting targets have defects is the key issue of the power system research.The rapid detection of the fittings and defective targets in aerial images of transmission lines based on computer vision is an effective way to solve the problem.This paper uses Feature Pyramid Networks(FPN)to achieve accurate detection of the fittings and their defect targets in aerial images of transmission lines.In the light of the transmission line inspection requires high accuracy in the detection of the fittings and their defect targets,the fittings and their defective targets have large scale changes in aerial images,and some of the fittings are small-scale targets,which lead to the problem of low detection accuracy.This paper presents a target detection method for the fittings and their defects in transmission lines based on improved FPN.First of all,the research group builds its own aerial photography fitting images data set,fine-tunes the parameters of the FPN model and explores the most effective hyperparameters.Then,for the problem that high-level semantic information and location detail information in the hierarchical depth feature cannot achieve good results at the same time.In this paper,the bilinear interpolation method is used to up-sampling for feature fusion,high-level semantic information is transferred to each feature level,and high-level semantic information is added to the feature map with detailed positioning information.Aiming at the hierarchical prediction of multi-scale targets and small-scale targets,a dual-parameter constraint method is proposed,and shape constraints are added to the resolution-based target region feature prediction layer selection method.Through experimental verification,the average detection rate of the above-mentioned improved method to FPN detection method reaches 82.37%,and the performance is improved by 3.25%.In order to obtain better target feature and effective expression of the target feature,this paper adopts the recursive feature pyramid,through the use of bottom-up feedback connection,to optimize the feature extraction network of the fittings features.At the same time,it is proposed to use Neural Architecture Search(NAS)to obtain the atrous rate of the atrous convolution to expand the receptive field,so that the convolution is more effective for multi-scale fittings feature extraction.Experimental results show that the combination of recursive feature pyramid and NAS search the atrous rate of atrous convolution to improve FPN.Solving the problem of low accuracy of fittings target detection to a certain extent.Among them,the performance index average precision(AP)value increased by 3.27%,and the highest detection accuracy rate reached 92.23%.This research has laid a good foundation for further fault diagnosis of fittings and realization of intelligent inspection.