CT-based data generation for foreign object detection on a single X-ray projection.

Area of Science:

• Industrial Imaging
• Artificial Intelligence
• Non-Destructive Testing

Background:

• X-ray imaging faces limitations in detecting internal defects due to feature superposition.
• Current deep learning methods require extensive annotated data, hindering industrial application with variable products.

Purpose of the Study:

• To develop a computationally efficient method for generating artificial single-view X-ray data.
• To reduce the reliance on large annotated datasets for defect detection.

Main Methods:

• A computed tomography (CT)-based approach to create artificial single-view X-ray data from a few physical scans.
• Algorithmic modification of CT volumes to generate diverse training examples.

Main Results:

• The generative model achieved high accuracy using data from a single CT-scanned object, comparable to tens of real-world samples.
• Significant reduction in required training data, improved defect detection coverage, and enhanced generalizability.

Conclusions:

• The proposed CT-based generative approach enables accurate real-time foreign object detection with minimal training data.
• This methodology enhances the viability of deep learning for industrial quality control, especially with high product variability.