Anomaly Detection

Anomaly detection aims to identify unusual patterns or outliers in data that deviate significantly from the norm, often indicating potential issues or interesting events. This can involve spotting suspicious activity, fraud, security threats, or equipment failures. The goal is to proactively detect these deviations, allowing for timely intervention and mitigation of potential problems. In addition, most machine learning models operate under the closed-world assumption, where it is presumed that the distribution of test data matches the training data. In real-world scenarios, this assumption often fails, as models may encounter outlier samples during testing that differ from the training distribution. To be reliable, a model must not only perform well in known contexts but also identify and reject anomalous inputs. Therefore, the task of anomaly detection is an important aspect of reliability and trustworthiness of the models. The broad field of anomaly detection encompasses various modalities and tasks, with each one aligned to a particular real-world application. In the anomaly detection track, our main focus is on visual anomaly detection. Given the wide range of potential anomalies and the challenges associated with obtaining anomalous samples, unsupervised anomaly detection is the most practical approach in this area. In this context, the model is given a training set of solely normal samples to learn their distribution. During inference, any sample that deviates from this established normal distribution, should be recognized as an anomaly. The figure below illustrates two main types of visual anomalies commonly found in industrial inspection scenarios. First, structural anomalies which include defects like dents or scratches that disrupt the uniformity of the image, making it easy to isolate the affected area. Second, logical anomalies which involve more intricate relations and extend beyond typical defects. In this type of anomalies, while no individual element of the image is flawed on its own, the overall composition of these elements and the underlying logical rule is violated.