Analysis of the Region of Interest According to CNN Structure in Hierarchical Pattern Surface Inspection Using CAM

• Published in: Materials Vol. 14; no. 9; p. 2095
• Main Authors: Moon, In Yong, Lee, Ho Won, Kim, Se-Jong, Oh, Young-Seok, Jung, Jaimyun, Kang, Seong-Hoon
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 21.04.2021; MDPI
• Subjects: Artificial intelligence; Artificial neural networks; class activation map; Classification; convolutional neural network; Decision analysis; Defects; Fault diagnosis; hierarchical pattern; Human error; Human performance; Inspection; Quality control; Raw materials; region of interest; Structural hierarchy; Surface defects; surface inspection; class activation map; hierarchical pattern; convolutional neural network; region of interest; surface inspection
• ISSN: 1996-1944; 1996-1944
• DOI: 10.3390/ma14092095

A convolutional neural network (CNN), which exhibits excellent performance in solving image-based problem, has been widely applied to various industrial problems. In general, the CNN model was applied to defect inspection on the surface of raw materials or final products, and its accuracy also showed better performance compared to human inspection. However, surfaces with heterogeneous and complex backgrounds have difficulties in separating defects region from the background, which is a typical challenge in this field. In this study, the CNN model was applied to detect surface defects on a hierarchical patterned surface, one of the representative complex background surfaces. In order to optimize the CNN structure, the change in inspection performance was analyzed according to the number of layers and kernel size of the model using evaluation metrics. In addition, the change of the CNN's decision criteria according to the change of the model structure was analyzed using a class activation map (CAM) technique, which can highlight the most important region recognized by the CNN in performing classification. As a result, we were able to accurately understand the classification manner of the CNN for the hierarchical pattern surface, and an accuracy of 93.7% was achieved using the optimized model.