Self-supervised anomaly detection in computer vision and beyond: A survey and outlook

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Anomaly detection (AD) is vital across cybersecurity, finance, and healthcare for identifying unusual patterns.
• Deep learning advancements have driven significant progress in AD methodologies.
• Self-supervised learning (SSL) has emerged as a powerful paradigm, yielding novel AD algorithms that surpass current state-of-the-art.

Purpose of the Study:

• To provide a comprehensive review of current self-supervised anomaly detection (SSAD) methodologies.
• To detail standard SSAD methods, analyzing their respective strengths and weaknesses.
• To compare the performance of various SSAD models against each other and other leading AD techniques.

Main Methods:

• Review of existing literature on self-supervised anomaly detection.
• Technical exposition of standard self-supervised anomaly detection algorithms.
• Comparative analysis of model performance using established benchmarks.

Main Results:

• Self-supervised learning methods demonstrate superior performance in anomaly detection compared to traditional approaches.
• A detailed comparison highlights the relative effectiveness and limitations of different self-supervised techniques.
• The review identifies key trends and performance benchmarks in the field.

Conclusions:

• Self-supervised learning represents a significant advancement in anomaly detection.
• Future research should focus on developing more efficient and effective SSAD algorithms.
• Integration with other fields like multi-modal learning offers promising avenues for future development.