Improvement of Anomaly Detection Accuracy by Reducing the Camera Gap Based on Image Conversion Using Ideal Pair Images

In recent years, due to the progress of machine learning, the classification accuracy of normal and anomaly classes by image processing has been improved. There are many situations where ”the environment of imaging for learning classifiers” and ”the environment of imaging in the inspection site” are different in the general manufacturing site. In particular, it is assumed that cameras may differ due to malfunctions or model changes. Images obtained from two different cameras may have the same visual appearance, however, differences in sensor and sensitivity cause differences in each pixel. This difference affects the classifier that judges between normal and anomaly using machine learning and hinders the achievement of accurate automatic visual inspection. To solve this problem, we propose an image conversion method that reduces the differences between two cameras. The image obtained by the camera for inspection is translated to the image obtained by the camera for learning of classifier. In the evaluation experiment, the proposed method achieved an accuracy of 97.0% in classifying normal and anomaly images. The proposed method improved accuracy by 3.8% compared to the previous method without image conversion (93.2%).