High-Reality Defect Image Generation Based on Fusion of Information in Eigen Space

Abstract

This paper describes the method of data augmentation for achieving accurate visual inspection by machine learning. Recently, the application of machine learning for accurate visual inspection has been expected. Generally, machine learning requires many training data. When applying machine learning to visual inspection, a lot of normal images and a lot of defect images are required for training. However, there is a shortage of defect images for machine learning because it is difficult to obtain many defect images from manufacturing factories. This is a critical issue for achieving accurate visual inspection. We propose a method that extracts normal and defect information from images by the principal component analysis and fuses the information in Eigen space to generate high-reality defect images. Moreover, when using these generated defect images for training classifier, the accuracy of normal/defective product discrimination achieved 96.3% for the capsule dataset. In the case of the pill dataset, the accuracy achieved 91.3%. For the aluminum plate dataset, the accuracy achieved 97.8%. These results show the higher accuracy than those of previous data augmentation methods. This shows generated defect images by the proposed method are effective for training a classifier.