Research On Recognition Of Bare Ground In Aerial Image Of UAV Inspection

As my country’s social economy continues to develop,the power transmission line network is developing rapidly,and the increasing power construction projects are increasing,there is likely to be bare ground in the power construction site.These bare ground is one of the hidden dangers that cause many power safety accidents.The reason is:(1)The exposed surface may collapse under rain washing,(2)When the relevant electric construction engineering machinery is constructed above the exposed surface,it is easy to enter the safe range of the high-voltage line,causing accidents such as breakdown and short-circuit.Therefore,it is necessary to find out the bare surface area of the electric power construction site during the inspection.Currently,the use of drones instead of manual inspections is becoming the main method of power inspections.For this reason,it is necessary to study the automatic recognition method of the bare surface in the aerial image of the drone inspection,and realize the high-precision and high-efficiency automatic identification of the bare surface in the aerial image of the drone inspection.This paper constructs a UAV inspection aerial image data set containing bare ground,uses four methods to construct a bare ground recognition model,and finds out a bare ground recognition model suitable for loading on a UAV airborne platform.The main contents of this paper are as follows:(1)Construct the data set of exposed surface image.The images collected by drone aerial photography are first normalized to unify the size,and then processed such as histogram equalization to highlight the image details and enhance the contrast.To amplify the image data set,The image is rotated,Gaussian processing,etc.Divide the image data set into a training set and a test set proportionally.(2)The Mask R-CNN method is used to construct the bare surface recognition model.The feature pyramid network used by the Mask R-CNN model can better extract the low-level and high-level feature information of the exposed surface,and can complete the identification of the exposed surface through the method of candidate regions and the use of fully connected layers.This article adopts the Mask R-CNN Research and analysis of the network model of the company,and explore its performance on the small sample data set.The experimental results show that under the small sample data set,the Mask R-CNN model can recognize the bare surface in the aerial image of the drone inspection The accuracy is58%,the weight parameter scale of the model is 255 M,and the recognition time efficiency is160 ms.(3)Extracting a single image feature method is used to construct a bare surface recognition model.Due to the obvious contour features of the exposed surface,and the texture of the exposed surface is different from the background,this article focuses on the HOG(Histogram of oriented gradients)feature descriptor describing the edge of the target contour and the LBP(Local Binary Pattern)describing the target texture.Feature descriptors are used to characterize the principle and calculation method of target feature information,and the performance of the model trained by combining the two features with the SVM(Support Vector Machine)classifier in the target recognition of bare ground is evaluated.Experimental results show that although the model for manually extracting image features can meet the requirements in terms of weight parameter scale and recognition efficiency,the model for extracting HOG features of bare ground has an accuracy of 74% on the test set for bare ground recognition,extracting bare ground LBP The feature model has an accuracy of65% on the test set for bare ground recognition.(4)After analyzing and studying the principles and calculation methods of HOG features and LBP features,as well as their respective advantages and disadvantages,this paper proposes a method of fusing the two features to identify bare ground,and by setting the magnitude difference of the weight coefficient,In this way,the original order of magnitude difference between the two features can be offset,and the advantages of the two features can be used.The experimental results show that the model trained by the feature fusion method has an accuracy of more than 80% in the recognition of the exposed surface of aerial images,and the weight parameter scale and recognition efficiency of the model can also meet the application requirements of the UAV airborne platform.In summary,under the small-scale data set,the HOG and LBP features proposed in this paper are used to identify bare ground after feature fusion according to their weights,which is more effective than deep learning models and models that manually extract single features.The method not only guarantees a certain degree of accuracy in identifying the bare surface,but also has the characteristics of real-time.At the same time,the training time and the scale of weight parameters are lower,which meets the requirements of being carried on the airborne platform of the UAV.The exploration in this article also provides a new idea for real-time identification of targets in aerial images during UAV inspections.