Research On Defect Detection Method Of Solar Silicon Wafer And Cell Wafer Based On Multiscale Features

Defects in solar cells will affect their photoelectric conversion efficiency,quality and service life.In the manufacturing process of solar cell raw materials,silicon wafers,dislocation defects may occur due to defects in raw materials or improper manual operations,and the silicon wafers with excessive dislocation density would be discarded due to low photoelectric conversion efficiency.In the subsequent manufacturing process,transportation,printing and other process links or manual errors may produce a variety of solar cell surface defects.These solar cell surface defects with different sizes,shapes,and colors will affect the solar cells’ power generation efficiency and the service life.Aiming at the problems that the dislocation defect is strongly similar to the lattice background,and the shape and color of the solar cell surface defects are various,this thesis studies from the perspective of feature extraction and fusion.The research methods of dislocation defect segmentation and surface defect classification are as follows:(1)In order to achieve segmentation of dislocation defect regions in silicon wafers with non-uniform random texture background,this thesis presents a novel strong robust multiscale feature saliency map(MFSM)to accurately segment dislocation defects.In order to highlight the dislocation area and weaken the background information,a parameteroptimized atmospheric scattering model(PASM)is used to enhance image contrast and retain dislocation defect area information.Next,a contour detection function is obtained by fusing multi-scale and multi-channel gradient features to contain all boundary intensity information of the enhanced image.The watershed transform is used to close the contour arc obtained by the contour detector,and a super-level contour tree is constructed,and all contours are indexed to obtain a multi-scale feature saliency map to achieve accurate segmentation of dislocation defect regions.Experimental results show that this method effectively enhances image defect information,achieves effective segmentation of dislocation defects,and has good adaptability and robustness to complex backgrounds.(2)In order to achieve the automatic classification of multi-class solar cell surface defects,and to solve the problem of weak defect feature extraction,this thesis proposes a neural network model(referred to as SCeption)based on multi-channel and multi-scale convolution feature fusion.The SCeption model combines cross-channel deep convolution,multi-scale convolution feature fusion structure and maxout,combined with the Dense Net structure,strengthens the model’s multi-scale feature expression ability and the model’s spectral invariance,stimulates competition between neurons in different spectral channels,improves the learning ability of the network,enhances the non-linear fitting ability of the model,and greatly improves the classification performance of the network on the multi-class solar cell surface defect dataset.The SCeption model is convolved separately on the three channels of the image,which removes the spatial correlation of image features.It also uses multi-scale convolution feature fusion to obtain large-scale features and small-scale features,respectively.Through adaptive weights and maxout fuse features,the model representation is stronger and the semantic information is more abundant.Experimental comparison results with several other mainstream feature extraction networks show that the classification accuracy of the SCeption model is better than those types of networks,with fewer model parameters and faster detection speed.