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

Overview

• X-ray imaging is widely used to detect defects in industrial products on a conveyor belt
• In-line detection requires highly accurate, robust, and fast algorithms
• Deep Convolutional Neural Networks (DCNNs) can satisfy these requirements with enough labeled data
• Collecting large labeled datasets is challenging, so methods of X-ray image generation are considered
• X-ray scattering is computationally expensive to simulate but can greatly affect the accuracy of generated images
• The goal is to quantitatively evaluate the effect of scattering on defect detection

Plain English Explanation

Manufacturers often use X-ray imaging to check their products for defects as they move along an assembly line. To do this quickly and reliably, they need very accurate and fast computer algorithms. Deep learning models like Convolutional Neural Networks (CNNs) can meet these requirements, but they need a lot of labeled training data, which is hard to collect.

Instead, the researchers tried generating artificial X-ray images. They found that accounting for X-ray scattering - when the rays bounce around inside the object - is important for making the generated images realistic. Simulating scattering is computationally expensive, so the researchers wanted to understand how much it really matters for detecting defects.

Technical Explanation

The researchers used Monte-Carlo simulation to generate X-ray scattering distributions, then trained CNN models on data with and without scattering. They applied the trained models to the same test datasets and compared their probability of detection performance, focusing on the size of the smallest detectable defect.

For a model problem of detecting defects in cylinders, they found that when trained on data without scattering, the CNNs could reliably detect defects larger than 1.3 mm. Including scattering in the training data improved performance by less than 5%. However, for cases with a large scattering-to-primary ratio (1 < SPR < 5), the difference in performance could reach 15%, affecting the ability to detect defects around 0.4 mm smaller.

Critical Analysis

The researchers acknowledged that excluding scattering has the biggest impact on detecting the smallest defects, and the difference decreases as defect size increases. They also noted that the scattering-to-primary ratio significantly affects detection performance, highlighting the need for accurate simulation of this effect.

One potential limitation is that the study focused on a single model problem of defect detection in cylinders. The generalizability of these findings to other industrial products and defect types could be explored further. Additionally, the researchers did not investigate the computational cost tradeoffs of including scattering simulation in the training data generation process.

Conclusion

This research quantifies the importance of accurately modeling X-ray scattering when generating synthetic training data for defect detection algorithms. While excluding scattering has the largest impact on the smallest detectable defects, it becomes less critical for larger defects. The scattering-to-primary ratio is a key factor in determining the required accuracy of the data generation process. These insights can help manufacturers optimize their X-ray inspection systems and deep learning models for industrial quality control.