Deep Anomaly Detection for Automotive Components by Oversampling

Abstract

Training of deep neural networks (DNNs) requires large amounts of data. However, the automotive components that are the subject of this research have an extreme lack of defective product data due to rapid model changes and a low defective product rate during the manufacturing process. Additionally, the anomaly areas are negligible. Data augmentation (DA), which increases data by image transformations, is a method for solving data deficiency. Particularly, a deep convolutional generative adversarial network (DCGAN) is frequently employed in the medical industry. DA is shown to have an effect on not small anomalies but on images that are accounted by the classification target for a large percentage of the total image.Therefore, in the present study, we find the DA method that fits objects with very small anomaly areas. We use a method that combines existing methods and DCGAN because DCGAN only cannot be used in cases where target images are few. After increasing the data to some extent using existing methods for only the anomaly area, we increase it further using DCGAN and then paste the completed defect to the component images.The classification performance using DA, which modifies size, shape, and the like, based on the inspector's experience, yielded a recall value of 76.9%, whereas the performance using DA and DCGAN yielded a slightly lower recall value of 65.4%.