Feature Encoding With Autoencoders for Weakly Supervised Anomaly Detection.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised anomaly detection utilizes limited labeled data and abundant unlabeled data.
• Current deep learning methods for anomaly detection often struggle with insufficient annotated anomaly samples.

Purpose of the Study:

• To propose a novel strategy for transforming input data into a more meaningful representation for anomaly detection.
• To enhance the performance of anomaly detection systems with limited labeled data.

Main Methods:

• Leveraging an autoencoder to encode input data.
• Utilizing three factors—hidden representation, reconstruction residual vector, and reconstruction error—as a new data representation.
• Proposing a novel network architecture to integrate these three factors.

Main Results:

• The proposed strategy significantly improves anomaly detection performance.
• Demonstrated superior performance over competitive methods in extensive experiments.

Conclusions:

• The novel feature encoding strategy effectively addresses the limitations of insufficient annotated anomaly samples in weakly supervised anomaly detection.
• The proposed method offers a robust approach for enhancing anomaly detection capabilities.