Quantifying the effect of X-ray scattering for data generation in real-time defect detection.

Area of Science:

• Industrial Nondestructive Testing
• Medical Imaging Physics
• Machine Learning for Quality Control

Background:

• X-ray imaging is vital for industrial defect detection, requiring accurate and fast algorithms.
• Deep Convolutional Neural Networks (DCNNs) offer high performance but need extensive labeled data.
• Generating realistic X-ray data is challenging, necessitating methods to simulate physical effects.

Purpose of the Study:

• To quantitatively assess the impact of X-ray scattering on defect detection accuracy.
• To evaluate how simulated scattering affects the performance of DCNNs in industrial inspection.
• To determine the influence of the scattering-to-primary ratio on detection capabilities.

Main Methods:

• Monte-Carlo simulations were employed to generate X-ray scattering distributions.
• DCNNs were trained on datasets with and without simulated scattering effects.
• Performance was compared using Probability of Detection (POD) curves, focusing on smallest detectable defect size.

Main Results:

• DCNNs trained without scattering detected defects >1.3 mm; incorporating scattering improved performance by <5%.
• In cases with a high scattering-to-primary ratio (1 < SPR < 5), performance differences reached 15% (~0.4 mm).
• The effect of excluding scattering was most pronounced for the smallest detectable defects.

Conclusions:

• Excluding X-ray scattering from training data significantly impacts the detection of small defects.
• The scattering-to-primary ratio is a critical factor influencing detection performance and data generation accuracy.
• Accurate simulation of scattering is essential for robust DCNN-based industrial X-ray inspection systems.