2016 Volume 55 Issue 3 Pages 348-354
In manufacturing process, automation of visual inspection is in high demand. Thanks to recent progress in imaging technologies and computer vision technologies, use of automated visual inspection by image data is increasing. In this paper, we describe our two methods for Visual Inspection Algorithm Contest as an introduction of our automated visual inspection technologies. The first method can learn decision boundary of quality determining without image data of defective parts by use of semi-supervised learning. It is difficult to prepare a large number of defective samples because defects occur only in rare cases. Thus, we developed a new algorithm which can learn without defective parts by semi-supervised anomaly detection. The second method can classify defective samples, even if a shape of non-defective parts has a large variance. Some products such as primary processing stage parts, its shape has a large variance. Thus, we modified “Z-score” calculation method to increase robustness. As a result, we achieved high accuracy in the contest.
