DRepT: Anomaly Detection Based on Transfer of Defect Representation with Transmittance Mask

Hiroki Kobayashi, Manabu Hashimoto

Abstract: Automated surface anomaly detection is important in industry. The goal with this task is to classify the query image and its pixel as either “Normal” or “Anomaly” with high precision for industrial products. Unlike general two-class classification tasks, we can only obtain normal images (i.e., anomaly images cannot be obtained) for training because anomalies rarely occur in well-controlled production environments such as modern manufacturing factories. We propose a transfer-based anomaly detection method called Defect Representation Transfer-based Anomaly Detection (DRepT). This method can generate natural anomaly images for the target domain (i.e., real factory with only normal images) using knowledge of anomalies obtained in the source domain (i.e., known datasets with normal and anomaly images). We first focused on an anomaly image of an industrial product consisting of two elements: “background texture” and “defect representation”. At the source domain, a background texture image is generated from an anomaly image with texture invariance (i.e., defects are only removed) by using generative adversarial networks, and defect representation is modeled as the difference between an anomaly image and its background texture image with the newly proposed Gaussian mixture transmittance mask. At the target domain, anomaly images are generated by giving defect representation to normal images, and a anomaly detection model is finally trained. Compared with a method without real defects of other domains and a method based on pasting of the defect images, DRepT achieved highly precision.