Research On Capacitive Panel Surface Defect Segmentation Method Based On Deep Learning

Various defects are hardly evitable to be produced on the workpiece surface during the production process because of many causes,and these surface defects often affect the workpiece quality.Therefore,the workpiece surface defect inspection has become an indispensable process for the product quality control in industrial production.With the development of computer vision technology,automatic surface defect inspection methods via machine vision have been widely used to replace traditional manual inspection.However,these vision-based methods still have some challenges to detect these defects with small sizes and various shapes,and they often suffer from the overfitting problem because of the lack of sufficient training data.This thesis focuses on the surface defect segmentation of the capacitive touch panel via deep learning and provides the algorithm support for surface defect detection system.The main contributions of this thesis are given as follows:1.This thesis builds a surface defect dataset for capacitive touch panel.Meanwhile,in order to address the problem of insufficient training data,this paper proposes a data augmentation strategy for the dataset of capacitive touch pane,and this strategy adopts image geometric transformation,image aliasing and generating adversarial network(GAN)to enhance the training data.The experimental results on the augmented dataset show that the proposed data augmentation strategy can effectively alleviate the over-fitting problem for the model training.2.This thesis firstly analyzes the characteristics of touch panel defects,and proposes a multi-scale feature fusion defect segmentation algorithm based on the DeeplabV3+network.In order to detect small-scale defects,this proposed method constructs a backbone which can preserve more fine-grained information and process large feature maps,and also improves the ASPP module to enhance the perception ability of multi-scale features.Meanwhile,this proposed method designs a feature fusion module to integrate the useful information of shallow feature maps with the high-level feature maps in the decoding stage to improve the segmentation accuracy.A combined loss function integrating the weighted cross entropy with the Dice Loss is proposed to solve the problem of class unbalance.The proposed d multi-scale feature fusion defect segmentation algorithm can significantly improve the accuracy of touch panel defects.3.In order to deal with the problem of insufficient training data in industrial applications,this thesis proposes a dual-branch defect segmentation network for limited samples via the idea of few-shot semantic segmentation.The independent features of supporting images and query images are extracted by the dual-branch feature extraction network and pyramid module,respectively,and the correlation information of these features is obtained by cosine similarity.Meanwhile,the self-guided module is introduced to process the feature maps and vectors to obtain defect information in support image.Finally,this method fuses the query features and support features in the prediction stage to segment the query image.The experimental results show that the proposed defect segmentation method can improve the segmentation accuracy.