MSFlow: Multiscale Flow-Based Framework for Unsupervised Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Modeling

Background:

• Unsupervised anomaly detection (UAD) is crucial for applications where only normal data is available for training.
• Existing methods struggle with varying anomaly sizes, impacting detection and localization precision.
• Normalizing flows offer a probabilistic approach for anomaly detection but face challenges with scale variations.

Purpose of the Study:

• To develop a robust unsupervised anomaly detection and localization framework that addresses the challenge of varying anomaly sizes.
• To enhance the performance of flow-based models in distinguishing and pinpointing anomalies without prior anomaly information.
• To introduce a novel multiscale approach for improved anomaly detection and localization accuracy.

Main Methods:

• Proposed a novel multiscale flow-based framework (MSFlow) utilizing asymmetrical parallel flows and a fusion flow.
• Implemented multiscale aggregation strategies tailored for both image-wise anomaly detection and pixel-wise localization.
• Leveraged the probabilistic nature of normalizing flows to assign low likelihoods to anomalous data points.

Main Results:

• MSFlow significantly outperforms existing methods across three anomaly detection datasets.
• Achieved state-of-the-art (SOTA) performance on the MVTec AD benchmark.
• Reached a detection AUORC of 99.7%, localization AUCROC of 98.8%, and PRO score of 97.1% on MVTec AD.

Conclusions:

• The proposed MSFlow framework effectively generalizes across different anomaly sizes for unsupervised anomaly detection and localization.
• MSFlow represents a significant advancement in flow-based anomaly detection, setting new performance benchmarks.
• The tailored multiscale aggregation strategies contribute to the framework's superior performance in both detection and localization tasks.