A Generic Semi-Supervised Deep Learning-Based Approach For Automated Surface Inspection

Automatic surface inspection is the key to control quality in industrial manufacturing process.Recently,a new automated surface detection method has generated in the field of deep learning.This method can automatically learn advanced features from training samples.This method is also robust and can detect defects on different types of surfaces.However,this deep learning-based approach often relies heavily on manpower to collect and tag training samples.In view of the existing automatic surface detection methods rely on a large number of labeled data and can't make full use of unlabeled data,this dissertation proposes two general methods based on semi-supervised deep learning technology.The contents are as follows:(1)This dissertation puts forward a semi-supervised deep learning method based on Mix Match.Mix Match is a semi-supervised learning algorithm that uses entropy minimization and mixes labeled and unlabeled data using Mixup.Based on Mix Match,this dissertation proposes a technology that uses Cutout data enhancement on the basis of Flip and Crop and a new convolutional neural network model based on residual network structure,which realizes automatic defect detection of images.Experiments were conducted on a public data set NEU and an industrial data set CCL.Compared with other deep learning methods,the proposed method has the best performance.In addition,A comparative analysis of the model performance under different labeled sample shows that this method can achieve better performance even with fewer training samples.(2)This dissertation also proposes a semi-supervised deep learning method based on improved Mix Match.For the first semi-supervised deep learning method based on Mix Match,the unsupervised loss function only uses the L2 loss function.Laine pointed out that the L2 loss function provides stricter constraints than the KL loss function,but Laine did not provide quantitative comparison data.This dissertation proposes an unsupervised loss function calculation method that combines the L2 loss function and the KL loss function in a certain ratio.Aiming at too much unlabeled data and too little labeled data,which may cause the model to overfit the labeled data,the TSA method is introduced.Experiments were conducted on a public data set NEU and an industrial data set CCL.The experimental results show that the algorithm combining the TSA method and the unsupervised loss function can obtain better classification results.