DiffusionAD: Norm-Guided One-Step Denoising Diffusion for Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Manufacturing

Background:

• Anomaly detection is crucial in industrial manufacturing.
• Existing generative models suffer from poor reconstruction quality, limiting performance.
• Novel approaches are needed to improve anomaly detection accuracy and efficiency.

Purpose of the Study:

• Introduce DiffusionAD, a novel anomaly detection pipeline.
• Address limitations of previous generative models in reconstruction quality and inference speed.
• Enhance the fidelity and applicability of anomaly detection in industrial settings.

Main Methods:

• Reformulate reconstruction using a diffusion model in a noise-to-norm paradigm.
• Employ a segmentation sub-network for pixel-level anomaly scoring.
• Introduce a rapid one-step denoising paradigm for accelerated inference.
• Propose a norm-guided paradigm to integrate multiple noise scales for improved reconstruction.

Main Results:

• DiffusionAD outperforms state-of-the-art approaches on four benchmarks.
• Achieves hundreds of times acceleration in inference speed with comparable reconstruction quality.
• Demonstrates effectiveness and broad applicability in industrial anomaly detection.
• Code available at https://github.com/HuiZhang0812/DiffusionAD.

Conclusions:

• DiffusionAD offers a significant advancement in generative-based anomaly detection.
• The proposed methods effectively address reconstruction quality and inference speed limitations.
• DiffusionAD shows strong potential for real-world industrial manufacturing applications.