Robust defect detection method for improving inspection process

Abstract

Automatic inspection machines for metal products can inspect products in large quantities and at a high speed while separating products without defects from defective products by using unified criteria. However, it is difficult to set accurate criteria to detect minute defects, and this results in misclassification. This study focuses on inspection techniques to improve the classification accuracy for an actual inspection process. There are differences in variables, such as raw materials, light color to capture images, and degree of degradation of light brightness, across different manufacturing plants. Previous studies utilized machine learning, image processing, and statistical methods. However, it is not clear how those different methods should be combined for improving the classification accuracy. Therefore, this study provides a comprehensive survey of an optimal combination of methods to increase robustness and improve classification accuracy. Furthermore, the results also suggest that a combination of methods can achieve classification accuracy and robustness exceeding those of previous methods.