科研成果详情

With the development of deep learning techniques, an increasing number of anomaly detection methods based on deep neural networks have been proposed during the last decade. Nevertheless, these methods often suffer from catastrophic forgetting when trained on continuously arriving data samples, as deep neural networks quickly forget the knowledge obtained from previous training while adjusting to learning new information. In this work, we propose a contrastive lifelong learning model for image anomaly detection. Rather than adopting CNN-based neural networks as in other anomaly detection approaches to learn representations from training samples, we propose a contrastive learning framework for anomaly detection in which Vision Transformer (VIT) is adopted for extracting promising representations. Two nonlinear structures (projector and predictor) are integrated into our model, which is helpful in improving the performance of anomaly detection. Moreover, a lifelong learning framework that contains teacher and student networks is deployed in our model, which is able to mitigate the problem of catastrophic forgetting in image anomaly detection. By leveraging both lifelong learning and contrastive learning frameworks, our model is able to progressively perform image anomaly detection where the problem of catastrophic forgetting can be greatly mitigated. We demonstrate the effectiveness of the proposed anomaly detection method by conducting experiments on multiple image data sets.