| In recent years, more and more computer vision technology are applied in power system.Unmanned aerial vehicle and computer vision technology have replaced human labor in electrical fault of transformer substation and high-pressure transmission lines inspection. In order to improve the automation and intelligence during inspection, it is necessary for machine to recognize electric device independently. Most images stem from our lives, containing a huge of information. How to obtain further information from images is a great contention at the moment.This paper, take the lightning arrester as my experiment subject to studies the feature extraction and object recognition, laying a good foundation of electric fault in high-press transmission lines inspection. Base on the visual studio 2010, adopt the Adaboost classifier algorithm and Perceptual hash algorithm to extraction features and recognize object. Design and realize transformer substation and high-pressure transmission lines inspection tracking system.The main work in this paper are as fellows:1.First of all, collect a number of lighting arrester images data by camera. Due to the real video of transformer substation and high-pressure transmission lines inspection is unpubilshed. After studying transformer substation and high-pressure transmission lines inspection video in TV news, imitate the action of robots and UAV in inspection to film some videos. Pick up the most qualified images and video.2.Preprocessing,include:1)Image noise removing.In the inspection,motion blur is caused by camera’s moving. Considerate Wiener filtering work best on motion blur moving.So select the Wiener filter.2)Color images histogram equalization. In consideration of almost high-pressure transmission are set up in the wild and inspection in day time, images color equilibration will caused by light intensity and taking pictures in different angle. Therefore, we need to equalization primary color images to improve visual verisimilitude as well as feature extraction. In the middle of experiments,Ⅰ found that equalize RGB images directly is inappropriately. Visual verisimilitude would obtain better result if transform RGB model into YCbCr model before equalization.3) Convert images to gray images to reduce computation complexity. As we all know, arrays formulations are ideally suited to the storage facilities of computers, gray image has one channel while color image has three channel.Computer address a image in a way like one pixel by one pixel, thus,reducing channels means reducing computation data.3.Features extraction.There are several kinds of features are most frequently used. Such as color,texture,geometry shape,spatial,corner and so on. To improve the accuracy of the Adaboost classifier, extract the LBP and HOG features of object instead of one feature.4.Recognition lighting arrester. Adaboost algorithm is introduced to recognize target in images.Using creatersamples.exe program of opencv to create positive and negative training samples,then extract the LBP and HOG features of object.The next is training several weak classifier until algorithm convergence.Result of experiment shows HOG+LBP features in Adaboost algorithm has rapid convergence rate and good stability. Classifier of one feature obtain accuracy of 54.65% on test image data recognition,while classifier of multi-features reach 63.84%.5.Object tracking. Design a new tracking system base on the main characteristic of Adaboost classifier and perceptual hash which avoid their underfunded.Results of experiments shows good performance of new tracking system.Perceptual hash tracking experiment shows that accuracy of lighting arrester is only 10%. Perceptual hash combine with Adaboost tracking system reach accuracy of 55%. |
PDF Full Text Request