Background: In the industrial manufacturing process, manually labeling enough datasets is time-consuming, which hinders the training and deployment of defect detection models. Therefore, automatic defect detection and its classification is the premise of industrial production quality. Objectives: The study mainly discusses about the detection of the Hemispherical Surface of the valve core by machine vision method. Methods: The paper put forward a novel semi-supervised algorithm to detect the Hemispherical Surface of the Valve Core. Under the condition of the lack of labeled datasets, the paper used labeled and unlabeled samples for model training. This thesis proposed, for the first time, using the Mean Teacher semisupervised learning framework and then making changes to the model; firstly, this paper proposed to use the Stochastic Weight Average (SWA) algorithm to update the weight of the teaching model to enhance this model’s generalization ability. Furthermore, in order to select reliable datasets and calculate the consistency loss, this study also proposed an Uncertainty Filter (UF) method. Thirdly, the selection of hard-ware equipment, since the hemispherical surface is anisotropic, ring light source is used, which can lit the surface from top to bottom. Results: Experimental results show that in two different conditions, the classification accuracy can raise. On one hand, under the condition of training with a small amount of labeled datasets, the proposed semi-supervised learning model can achieve a classification accuracy of 90.51%; whereas, under the condition of the semi-supervised learning mechanism and a large amount of unlabeled datasets, the accuracy increases from 93.7% to 98.1%. Conclusion: This paper uses hemispherical metal surface as the dataset for the first time, and also innovatively optimizes the semi-supervised model. On the other hand, experimental comparative analysis indicates that the model proposed in this paper is significantly better than the comparison model, which lays the basic position for the defect detection of the hemispherical surface’s metal. At the same time, the novel semi-supervised algorithm can also be used to detect other metal part’s hemispherical surfaces.
Read full abstract