Patch Core++ : Augmenting Coreset for Superior Industrial Anomaly Detection

This paper proposes a novel method for effectively detecting anomalies on the surfaces of objects produced during industrial activities. The most effective way to detect anomalies reliably is to clearly identify each anomaly type, collect corresponding data, and train models specific to each type, which can potentially yield the best performance. However, in large-scale industrial environments, the types of anomalies are often highly diverse and difficult to define clearly. Consequently, collecting data for all types and training models accordingly is a highly challenging task. To address this issue, recent research has focused on methods that utilize only normal data, which is relatively easier to construct. The key to this approach is extracting various features from normal data that could occur in real-world scenarios and using them for model training. It is particularly important to ensure that the model remains robust under various conditions, such as noise, brightness, and contrast variations that may arise during image capture. To achieve this goal, numerous studies have applied various augmentation techniques to train models on the diverse conditions that can occur in normal states. However, when augmentation is applied at the image level, it necessitates reprocessing the image before feeding it back into the model, leading to increased computational costs proportional to the image size. Therefore, this paper proposes a method to obtain a noise-robust normal Coreset by applying a feature level augmentation to the Patch Core model. This method extracts common features between normal data and noise-augmented normal data. Finally, the utility of the proposed method is validated on the BeanTech Anomaly Detection dataset through various performance metrics. Notably, the fact that INSTANCE AUROC improved by approximately 2.6 for Product 2, which was particularly weak, is a remarkable result.