Classification System Of Polycrystalline Silicon Solar Panels Based On Image Texture Features

With the increasing shortage of energy and the severe air pollution,the work of new sources,such as solar energy,wind energy,nuclear energy and so on,is pushing forward.Solar energy is usually cultivated via photovoltaic technology.The solar panel is the carrier to solar power,whose quality determines the photoelectric conversion efficiency.The quality check of solar panels concentrate on the defects of solar panels’ surface.Traditional methods pay more attention on size defects or gate defects,while the grain boundary defects are less of a concern.In this paper,based on the characteristics of grain boundary defects and the industry require,“Dark Crystal”,“Light Crystal” and “No Crystal” solar panels have been classified,which makes it possible to automate the identification of grain boundary defects.Specific research contents and achievements are as follows:System design.Two serial BP neural networks,used as classifiers,complete the classification of solar panels based on grain boundary characteristics.The first one uses U2 LBP feature to distinguish between "With Crystal" and "No Crystal" images,while the second one uses the Local Contrast Feature to find out "Dark Crystal” and "Light Crystal" images.Image preprocessing.With the help of Hough Transform and Image Normalization,all gate lines will be erect.Afterwards,gate lines will be eliminated or blurred on the corresponding images of different types,which can improve the recognition performance efficiently.Feature extraction.This paper recalls the calculation of LBP algorithm,analyzes the advantages of U2 LBP features,and apply U2 LBP method on images.Then,the hypothesis test will test U2 LBP features sequentially,and eight features with significant differences will be signed.With these eight features,the first classification will recognize whether the image has crystal patterns or not.Finally,the "With Crystal" image will be divided into 32 pieces,and the local contrast of each piece will be calculated.The classification on "Dark Crystal" and "Light Crystal" will complete,attribute to the degree of contrast.System Performance.Three groups of controlled experiments are designed,which are used for analyzing the changes of the recognition rate when the system run in different numbers of training samples or different sample-features,and also analyzing the changes of the running time in different operating systems and different hardware environments.The experimental results show that the recognition rate of the classification system is above 90% and the recognition time is less than 900 milliseconds per piece,which meets the speed requirements of 4000 pieces per hour for the industrial production line.