Voxel-wise segmentation for porosity investigation of additive manufactured parts with 3D unsupervised and (deeply) supervised neural networks

• DOI: 10.1007/s10489-024-05647-z
• Publication Date: 2024-10-31T09:03:52Z
• Authors: Domenico Iuso; Soumick Chatterjee; Sven Cornelissen; Dries Verhees; Jan De Beenhouwer; Jan Sijbers
• Keywords: Computer. Automation; Physics

Abstract Additive Manufacturing (AM) has emerged as a manufacturing process that allows the direct production of samples from digital models. To ensure quality standards are met in all batch, X-ray computed tomography (X-CT) is often used combination with automated anomaly detection. For latter, deep learning (DL) detection techniques increasingly used, they can be trained to robust material being analysed and resilient poor image quality. Unfortunately, most recent popular DL models have been developed for 2D processing, thereby disregarding valuable volumetric information. Additionally, there notable absence comparisons between supervised unsupervised voxel-wise pore segmentation tasks. This study revisits (UNet, UNet++, UNet 3+, MSS-UNet, ACC-UNet) (VAE, ceVAE, gmVAE, vqVAE, RV-VAE) porosity analysis AM X-CT images extends them accept 3D input data 3D-patch approach lower computational requirements, improved efficiency generalisability. The were using Focal Tversky loss address class imbalance arises low training datasets. output was post-processed reduce misclassifications caused by their inability adequately represent object surface. findings cross-validated 5-fold fashion include: performance benchmark models, an evaluation post-processing algorithm, effect In final on test set quality, best performing model UNet++ average precision 0.751 ± 0.030, while ceVAE 0.830 0.003. Notably, model, its technique, exhibited superior capabilities, endorsing preferred task.