Detection Of Surface Defects On Solar Cells Based On Deep Learning

Crystalline silicon solar cells can convert light energy into electric energy,and play an important role in the development of new energy.Production process defects or human operation errors will lead to brokencell,cracks,virtual welding and other minor defects on the surface of the solar cell,which will reduce the working efficiency of the solar cell components and reduce the service life of the solar cell.Because of the subtle and unobservable characteristics of surface defects,there are many difficulties in the detection of solar cell surface defects.With the rapid development of artificial intelligence and machine vision,the deep learning method based on convolutional neural network model is applied to the detection of solar cell surface defects,and has achieved better detection performance,which is a new research hotspot in this field.This paper focuses on researching the improvement of deep learning network model used in detecting solar cell surface defects,aiming to further improve the accuracy of detection and reduce the error of defect location regression.The main work of this paper includes the following two aspects:First of all,in order to solve the problems such as high rates of error detection and missing detection,poor position accuracy and so on,This paper proposes an algorithm for surface defect detection of solar cell by fusing multichannel convolutional neural networks.A new fusion scheme of detection candidate box is designed.The Faster R-CNN network based on VGG-19 and the full convolution R-FCN network model based on Resnet-101 are complementary fused,which effectively improves the precision of defect location regression,and reduces the rate of missed detection and false detection.Secondly,to solve the problem of low feature extraction capability for Faster R-CNN to detect the solar cell surface defect,this paper introduces the cross layer connection structure,and uses the multi-scale feature extraction mechanism to extract the target candidate box,so that the network does notneglect the shallow layer information while learning the deep layer feature information,and increases the accuracy of the candidate box prediction.The improved network structure can better adapt to the scale change of the narrow and long surface of solar cell,and effectively reduce the error rate of detection.