Classification And Detection Of Near-infrared Image Defects In Photovoltaic Solar Cells Based On Supervised Learning

Photovoltaic solar cells are important energy supply components for many military and civilian products(such as satellites,space stations,photovoltaic power stations,etc).However,the defects of photovoltaic cells will have a negative impact on its product quality,such as reducing the power generation efficiency and service life of photovoltaic products.Therefore,the automatic detection of defects in photovoltaic cells has great significance and value.Photovoltaic cells are accompanied by defects in the process of manufacturing and packaging power generation.The existing surface detection algorithms still have shortcomings in terms of defect detection accuracy and anti-interference ability.Therefore,this thesis manually extracts image features from artificial feature descriptors,transition to convolutional neural network to automatically extract features,and proposes three defect recognition methods for the defects of photovoltaic cells in different scenarios.The contributions and contents of this thesis are as follows:1)Aiming at the problem of defect classification under the complex background interference of photovoltaic cells,this thesis proposes a Center Pixel Information Center Symmetric Local Binary Pattern(CPISC-LBP),which describes the image gradient texture features.CPISC-LBP LBP integrates the gradient features of the central pixel into the central symmetric binary mode in a thresholding manner,making fully use of the spatial correlation between the central pixel information and the surrounding pixel information to make the extracted defect features more robust.In addition,in order to capture the global robustness of the image,a bag of features model(Bag of CPICS-LBP,BCPICS-LBP)based on unsupervised feature extraction is proposed.This model uses clustering to perform similarity feature analysis to obtain the global features of the image,and finally applies the classifier for defect classification.The experimental results show that the two feature extraction methods in this thesis have achieved excellent performance.2)Aiming at the problem of original image defect detection under the background interference of photovoltaic cells,this thesis proposes an Faster RPAN-CNN,an end-to-end defect detection framework based on Region Proposal Attention Network(RPAN).First,we connect the proposed channel attention and spatial attention to form a complementary relationship to obtain the proposed complementary attention network(CAN),and then integrate the complementary attention network into the Region Proposal Network to obtain the proposed Region Proposal Attention Network(RPAN).Because CAN has the function of suppressing complex backgrounds and highlighting target defect areas,RPAN can more accurately recommend candidate regions frames containing defects.Experimental results show that the proposed Faster RPAN-CNN detection model achieves a very high improvement in defect detection accuracy.3)Aiming at the problem of small target defect detection of photovoltaic panels under complex backgrounds,this thesis proposes a deep learning target detection framework AG-YOLO v3(Attention Guided You Only Look Once Vision 3).AG-YOLO v3 uses the proposed Self-Attention Gate Stackable Network(SAGSN)to capture defect information between feature layers of different scales.SAGSN first uses the proposed Self-Attention Gate(SAG)network to achieve multi-scale fusion of high-level features and low-level features,and suppresses complex background features,and then fuses the output features of the Self-Attention Gate Network with high-level features,which can make fully use of high-level semantic features and improve the feature expression ability of small target defects.The experimental results show that the AG-YOLO v3 proposed in this thesis achieves excellent small target defect detection results.