Improving Defect Detection in GDXray Castings via Inverse Problem-Based Deep Learning

Baghaeian, Mansour; Abbasi, Ehsan

doi:10.22068/ase.2026.724

Volume 16, Issue 1 (3-2026) ASE 2026, 16(1): 4922-4936 | Back to browse issues page

‎ 10.22068/ase.2026.724

Mendeley

Zotero

RefWorks

Baghaeian M, Abbasi E. Improving Defect Detection in GDXray Castings via Inverse Problem-Based Deep Learning. ASE 2026; 16 (1) :4922-4936
URL: http://ase.iust.ac.ir/article-1-724-en.html

Improving Defect Detection in GDXray Castings via Inverse Problem-Based Deep Learning

Mansour Baghaeian

, Ehsan Abbasi

Assistant Professor, Department of Mechanical Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran

Abstract: (57 Views)

In metal casting, detecting defects like pores and cracks in X-ray images is crucial for product quality and safety. This study presents an advanced U-Net architecture for semantic segmentation of defects in the GDXray dataset, achieving superior accuracy. By formulating defect detection as an inverse problem reconstructing material density from X-ray projections the method integrates transfer learning, data augmentation, and Convolutional Block Attention Modules (CBAM) to address low contrast-to-noise ratios and limited data. Pretrained on synthetic Radon transform projections, the U-Net, enhanced with CBAM, sharpens focus on defect regions, improving boundary precision by 5%. Data augmentation, including rotations, flips, and noise injection, generates 5,000 synthetic images to overcome data scarcity. Experiments on 2,727 grayscale GDXray images demonstrate a mean Intersection over :union: (mIoU) of 0.85, a 15% improvement over baseline U-Net models, with 97.8% accuracy for pores and 94.5% for cracks. The inverse problem approach reduces false negatives by 12%, excelling in noisy conditions. Compared to methods like Mask R-CNN, this approach advances non-destructive evaluation (NDE) for casting applications, ensuring reliability and safety. Validated on laboratory X-ray data, the model offers a scalable solution for industrial defect detection. Future work will optimize computational efficiency and explore multi-modal data to enhance robustness.