Steel Surface Defect Recognition Based On Graph Neural Network

Steel is an indispensable raw material in modern industrial production.In the production process,different types of surface defects would affect the application of steel products,so it is important to recognize the defects correctly to ensure product quality and reduce losses.Traditional recognition mainly relies on human labor,which relies too much on expert knowledge and the fatigue of the personnel.In recent years,with the advancement of computer technology and the introduction of deep learning algorithms such as Convolutional Neural Networks(CNN),more and more automatic inspection systems have been proposed and applied to real-world production.However,most of the current deep learning algorithms ignore interclass similarities and intra-class variations existing in steel surface defects,which will affect the performance of these algorithms on application.Therefore,this paper proposes to use a graph-based model to solve this problem;meanwhile,the graph-based model is improved to adapt to solve the semi-supervised learning,class imbalance and small samples.Firstly,a deep learning framework based on graph model is proposed to solve the interclass similarities and the intra-class variations.In order to construct an inductive graph-based algorithm,this research uses convolutional neural network to extract features of defect samples,and then constructs a micrograph through similarity calculation.By using the characteristics of graph convolutional network,the inter-class distance is reduced and the intra-class distance is reduced.Experiments verify that the framework containing graph convolutional neural networks has better feature distribution than the original method,even the untrained models have clear feature clusters.Better feature distribution also helps the proposed framework have better performance.Secondly,a graph-based semi-supervised learning framework is proposed.This research divides the training process into training and labeling state.The model is firstly trained on the labeled dataset.When the model accuracy reaches a predefined threshold,the parameters are fixed and the model is used to predict the unlabeled data and merge it into the labeled dataset.Then,the model continues to train on the new labeled dataset to get the final model.Experiments verify that the proposed framework can solve the inter-class similarities and the intra-class variations while making full use of the information of unlabeled samples,so it has better feature distribution than traditional semi-supervised algorithms.Even when there are fewer labeled samples,it can also have better performance.Thirdly,an anchor-based class-balanced graph convolutional network is proposed to address the coupling between inter-class similarities & intra-class variations and class imbalance.This method proposes class-balanced graph that can solve the impact of class imbalance on information propagation,and then proposes anchor vectors to reduce the impact of information imbalance on graph construction,so that the constructed network graph can present the relations between samples correctly.Experiments verify that this method can solve the coupling between inter-class similarities & intra-class variations and class imbalance,so it has better feature distribution and better performance on difficult-to-identify defect classes,and has the best performance compared to traditional algorithms.Fourthly,a graph-guided convolutional neural network is proposed to improve the performance with small samples.In order to reduce sample dependence,this method uses the relationship between samples to define a graph,which defines a loss function to realizes the guidance of CNN and improve CNN's ability to resolve inter-class similarities and intra-class variations.Experiments verify that the graph guidance can improve the feature extraction of CNN.Therefore,compared with the original model,the algorithm model with graph guidance has better feature space distribution and therefore better performance.Finally,the full thesis is summarized and the future research direction is prospected.