Hardware free industrial anomaly detection using physics informed thermal proxies from RGB

Area of Science:

• Computer Vision
• Materials Science
• Industrial Monitoring

Background:

• Thermal infrared imaging is ideal for industrial anomaly detection but expensive.
• High cost of calibrated thermal cameras limits accessibility in industrial settings.

Purpose of the Study:

• To develop a cost-effective alternative to thermal cameras for industrial anomaly detection.
• To validate a synthetic thermal proxy derived from RGB input.

Main Methods:

• Utilized Kirchhoff's law of thermal radiation and Lambertian surface assumptions to create a synthetic thermal proxy (ε = 1 - L) from RGB images.
• Combined the synthetic thermal proxy with frozen CLIP ViT-B/16 patch features and PDE-based multiscale residuals.
• Integrated a physics-informed Laplacian residual for enhanced detection on repetitive textures.

Main Results:

• Validated the synthetic thermal proxy against lock-in thermography, achieving a mean Pearson r of 0.4805 (r > 0.78 for Lambertian surfaces).
• ThermalCLIP achieved a mean image-level AUROC of 0.9441 across 15 MVTec AD categories.
• Demonstrated strong zero-shot transfer capability to the VisA dataset with an AUROC of 0.8724.

Conclusions:

• The proposed synthetic thermal proxy and ThermalCLIP model offer a viable, low-cost solution for industrial anomaly detection.
• Physics-informed residuals enhance the discriminative power of appearance-based methods, particularly for textured surfaces.
• ThermalCLIP shows significant potential for broad industrial application and zero-shot anomaly detection tasks.