M3DM-NR: RGB-3D Noisy-Resistant Industrial Anomaly Detection via Multimodal Denoising
Summary
This study introduces a novel framework for industrial anomaly detection using both RGB and 3D data, even with noisy datasets. The M3DM-NR method effectively filters noise and improves detection accuracy by leveraging multi-modal features.
Area of Science:
- Computer Vision
- Machine Learning
- Industrial Automation
Background:
- Current industrial anomaly detection relies heavily on unsupervised learning with clean RGB images.
- Real-world datasets often contain noise, and both RGB and 3D data are vital for robust detection.
Purpose of the Study:
- To develop a noise-resistant framework for RGB-3D multi-modal anomaly detection in industrial settings.
- To address the challenge of noisy datasets in practical anomaly detection scenarios.
Main Methods:
- Proposed the novel noise-resistant M3DM-NR framework utilizing CLIP's multi-modal discriminative capabilities.
- Implemented a three-stage approach: Suspected References Selection, sample denoising via intra-modal comparison and multi-scale aggregation, and Point Feature Alignment for final detection.
Main Results:
- The M3DM-NR framework demonstrated superior performance in RGB-3D multi-modal noisy anomaly detection.
- Achieved state-of-the-art results, outperforming existing methods in extensive experiments.
Conclusions:
- The M3DM-NR framework effectively handles noisy datasets in multi-modal anomaly detection.
- The proposed method offers a robust solution for industrial anomaly detection by integrating RGB and 3D data.
