Diverse Anomaly Feature Distributions Learning for Open-Set Supervised Anomaly Detection

Open-set supervised anomaly detection (OSAD) uses a small number of samples from the seen anomaly category during training, enabling the model to detect both seen anomalies and unseen anomaly types that are not present during the training phase. However, existing OSAD methods typically treat all anomaly categories as belonging to the same distribution. They do not learn the diverse distributions of anomalies, which limits their generalization ability to unseen anomaly categories. To address this issue, we propose a novel method called Diverse Anomaly Feature Distributions Learning (DAFDL). DAFDL is built on the CLIP model and consists of an image head and a structure head, each designed with different textual descriptions. The image head is responsible for learning the distributional differences of the overall image features and shares the textual information with the structure head during joint training. The structure head introduces a hypergraph to learn the feature differences caused by structural variations in anomalous samples, while leveraging the shared textual information to guide the hypergraph to focus on anomalous samples. DAFDL treats anomaly categories with common representations in structurally changed regions as belonging to the same distribution, enabling the learning of diverse anomaly distributions and enhancing the model's capability to recognize unseen anomalies. Experiments on four anomaly detection datasets show that DAFDL significantly outperforms existing mainstream OSAD methods in detecting both seen and unseen anomalies.