Deep Learning for Robust Defect Detection in Industrial X-Ray Images with Minimal Real-Data Dependency

Defect detection in aluminum casting components via X-ray imaging is essential for quality assurance but is hindered by challenges like low contrast and diverse defect geometries. Traditional automated defect recognition (ADR) methods often lack robustness across varying datasets. This study introduces a convolutional neural network (CNN) framework that combines simulated and real X-ray images to address these limitations. A simulation pipeline generates realistic 2D radiographs by embedding segmented 3D defect models, obtained from industrial CT scans, into arbitrary 3D objects, followed by 2D projection. This framework incorporates a carefully curated real-image dataset with annotations to overcome the gap between synthetic and real data. The model is built on a U-Net architecture optimized for single-channel grayscale images, using bilinear up-sampling to reduce overfitting while improving generalization. A tiling-based approach enables compatibility with arbitrarily shaped images, enhancing adaptability across datasets. Additionally, transfer learning minimizes the need for extensive real-world data and allows efficient retraining for new scenarios. Despite annotation inaccuracies affecting conventional metrics like precision and recall, qualitative evaluation highlights the model's effectiveness in detecting subtle defects. Its scalability and ability to infer from diverse X-ray images position it as a transformative tool for automated defect detection, reducing reliance on extensive real data and improving industrial quality control.