Improving automatic defect recognition on GDXRay castings dataset by introducing GenAI synthetic training data Articles uri icon

publication date

  • April 2025

start page

  • 1

end page

  • 13

volume

  • 151

International Standard Serial Number (ISSN)

  • 0963-8695

Electronic International Standard Serial Number (EISSN)

  • 1879-1174

abstract

  • X-rays are a Non Destructive Testing (NDT) technique commonly employed by aerospace, automotive or nuclear industries when the structural integrity of some parts needs to be guaranteed. Industrial dataset are now available with the introduction of Digital Radiography (DR) X-ray machine and are the basis for Automated Defect Recognition (ADR) systems based on Neural Network (NN) object detection models. However, building a big enough dataset is not easy and takes a long time in a production environment, delaying the introduction of ADR models. A potential solution is to use Generative Artificial Intelligence (GenAI) to synthesise new images. However, these models fail to generate full realistic images due to the subtle nature of X-ray images. Hence, this paper propose a combination of flawless images and synthetic defects generated by a novel Scalable Conditional Wasserstein GAN (SCWGAN) model. Such synthetic defects are introduced in the target images by a location algorithm that uses a mask image defining the allowable defective areas, the expected Gaussian or Poisson noise level and the defect size and aspect ratio. By creating such synthetic dataset and combine it with the original GDXRay dataset, our proposed detection system achieves an improvement of 17 % in mAP@IoU=0.5:0.95 (our target metric to reduced uncertainty on defect location) with regards the baseline model trained with only real images. As a secondary metric, to allow comparison with other studies, the model also achieves 96.0 % mAP@IoU=0.50, which exceeds the maximum accuracy available on current literature for the evaluated dataset.

subjects

  • Computer Science

keywords

  • x-rays; castings defects; automated inspection; adr; wgan; neural network; generative ai