ActiveInspect: GRPO-Optimized Multi-Sensor Evidence Selection for Industrial Defect Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Automated visual inspection is crucial for manufacturing quality assurance.
• Current methods often use fixed observations, limiting detection system effectiveness.
• Vision-language models (VLMs) and reinforcement learning (RL) are applied but constrained by observation sets.

Purpose of the Study:

• To introduce ActiveInspect, a novel framework for budget-constrained sequential evidence selection in automated visual inspection.
• To improve the accuracy-observation trade-off in industrial defect detection.
• To enable intelligent selection of views and modalities for enhanced defect identification.

Main Methods:

• Formulated inspection as sequential selection of multi-view, multi-modal evidence under a budget constraint.
• Initialized policy with perception-activated supervised fine-tuning (PA-SFT) and optimized with group relative policy optimization (GRPO).
• Converted depth and point-cloud data to VLM-compatible renderings; used structured memory for evidence integration.

Main Results:

• ActiveInspect demonstrated consistent improvements in the accuracy-observation trade-off across multiple datasets (Real-IAD D³, MVTec 3D-AD, etc.).
• On Real-IAD D³, it increased image-level area under the receiver operating characteristic curve (I-AUROC) from 0.890 to 0.906 while reducing observations.
• Achieved near-exhaustive performance with significantly fewer observations and reduced inference time, especially for geometry-dependent defects.

Conclusions:

• ActiveInspect offers a more efficient and accurate approach to automated visual inspection.
• The active selection of evidence significantly enhances defect detection capabilities.
• This method is particularly beneficial for identifying complex, geometry-dependent defects in manufacturing.