Research On Industrial Defect Detection Based On Deep Learning

Compared with the artificial defect detection method,the industrial defect detection method based on machine vision not only saves labor cost but also has high efficiency and flexibility,and has become one of the main research directions of automatic detection system.At present,machine vision-based defect detection algorithms mainly rely on some rule restrictions and some hand-designed features.The defect detection algorithms in this mode cannot adapt to the diversity of defects and are not robust.Therefore,this thesis takes the defect detection problem of polycrystalline silicon solar cell as the object,and exploits the data driving and feature automatic learning advantages of deep convolutional neural network to develop a research on industrial defect detection based on deep learning.The specific research work is as follows:Firstly,for the defect detection of polycrystalline silicon wafers,a defect segmentation method based on Convolutional Neural Network is proposed.The background of the polycrystalline silicon wafers is complicated,the position and shape of the defect are highly random,and generally the defect area is small.In order to simultaneously segment defects of different sizes and reduce the impact of defect pixels and target pixels distribution imbalance on segmentation accuracy,the algorithm first uses Region Proposal Networks(RPN)to extract all possible defect regions.Then,the recommended area obtained by the RPN is divided into image blocks of a certain size,and be sent to the segmentation network to obtain the final segmentation result.Since deeper convolutional neural networks lose information on smaller target defects,the segmentation network of this method uses a U-net segmentation network improved by dilated convolution.Compared with other segmentation algorithms,the experimental results show that the proposed segmentation algorithm has great advantages in detecting multiple defect types.Secondly,a method for detecting broken gate defects based on Densely Connected Convolutional Networks(DenseNet)is proposed.Broken gate defects have the characteristics of complex background,low contrast and weak semantic information,so the low-level texture features are very important for detection.In order to better learn the low-level texture features,with a limited amount of data,this method first improved a lightweight densely connected network for feature extraction based on DenseNet structure.These features are then entered into the RPN network and the location-sensitive score layer respectively.The RPN network performs the detection of the defect area,and the position sensitive score layer calculates the position probability of the defect.Finally,through the regional pooling layer,the defect area information extracted by the RPN is mapped onto the position sensitive score map to obtain the detection result.This method makes use of the texture features and replaces the fully connection layer with a fully convolution layer,which improves the detection speed while achieving reliability detection.