A Study And Application On Few Shot Learning-Based Defect Visual Recognition For TFT-LCD

In the production line of TFT-LCD(Thin Film Transistor-Liquid Crystal Display),the visual recognition of defect is of great significance since the visual recognition can guide the further improvement of process.Among the various types of defect,Mura defect is the most important one since that it may cause severe client performance issue and meanwhile it is difficult to recognize the type of Mura defect.Due to the high demand for labeled data,deep learning-based methods can not be employed in production line.And for improving the practical value of deep learning-based methods and deal with the high demand of labeled data,we proposed active learning-based method and transfer learning-based method in this study.Details are as follows:(1)For the lack of labeled data in real industrial application and the difficulty for human labeling process,we proposed an active learning-based Mura classification method.The most valuable unlabeled data can be chosen to be labeled by human through active learning.Moreover,we proposed a similarity-measurement-based module for auto-labeling before human labeling,which can further decrease the demand of human annotation.Through the method,only a small amount of human labeling effort is needed and most of unlabeled data are labeled automatically by algorithm and hence reduce the dependence for human labeling effort.And the result shows that proposed method can reduce almost 50% of human annotation than existing deep learning methods.(2)For the model degradation when a model tested in another similar task,we proposed a transfer-learning-based method.Through the pixel-level transfer and feature-level transfer,the model can keep discriminative capability as possible.Though the algorithm,the knowledge of existing high-performance system which has been employed in real application can be transferred to the new industrial scene without newly human labeled data.And it can get rid of human labeling effort as much as possible.In experiment of dataset,the result shows that the performance loss is only 5.2%,achieving state-of-the-art compared to CyCADA(Cycle-consistent adversarial domain adaptation),DAAN(Domain-Adversarial Training of Neural Networks)and ADDA(Adversarial discriminative domain adaptation).(3)For the inspection of scratch in mobile phone covers,existing methods that are based on two-dimension technology may be easy to be affected by texture and noise.We proposed a photometric stereo-based method for the scratch inspection.Through fixing the camera and object and varying illumination direction,the normal of object surface can be calculated,then the shape of scratch can be reconstructed.The method can easily detect the scratch from the interruption of noise and texture.Comparing to traditional twodimensional-based methods,proposed method can promote robust and accuracy greatly.And the result shows the mean angle error is 12.045,which is better than Entropy Minimization and Diffuse Maxima method.As for the real application in production line.We also design the corresponding AOI equipment which can match the demanding of production line.Moreover,we also employ the proposed equipment and algorithms in display panel companies.In application of Mura inspection,the proposed active learning-based method can reduce 50% of human annotation effort comparing to existing deep learning methods.And the proposed transfer-learningbased method can achieve 90.1% in accuracy without newly human annotation.In the application of scratch inspection,the proposed photometric stereo method achieve 95% performance in detection rate and 9.1% in over-detection rate.The results show that proposed system is efficient and very practical.