Analysis of AI-Based Single-View 3D Reconstruction Methods for an Industrial Application.

Area of Science:

• Smart Manufacturing
• Computer Vision
• Machine Learning

Background:

• Machine learning (ML) provides crucial insights in smart manufacturing without deep domain expertise.
• 3D reconstruction from 2D images offers advantages over 3D scans for quality control, including simpler setups and faster data acquisition.
• Existing methods for 3D reconstruction from single 2D images are challenging and have seen various proposals over decades.

Purpose of the Study:

• To investigate, evaluate, and compare deep learning algorithms for 3D reconstruction from single 2D grayscale images of laser-welded components.
• To assess the suitability of stacked autoencoders (SAE), generative adversarial networks (GANs), and U-Nets for this task.
• To introduce and evaluate the stacked dilated U-Net (SDU-Net) for 3D reconstruction.

Main Methods:

• Comparison of three deep learning architectures: Stacked Autoencoder (SAE), Generative Adversarial Networks (GANs), and U-Nets.
• Evaluation of different GAN variants, identifying Wasserstein GANs (WGANs) as the most robust.
• Introduction and application of the Stacked Dilated U-Net (SDU-Net) utilizing stacked dilated convolutions.

Main Results:

• Wasserstein GANs (WGANs) demonstrated robustness among tested GAN variants.
• The U-Net architecture, typically used for semantic segmentation, was applied to 3D reconstruction, achieving notable results.
• The Stacked Dilated U-Net (SDU-Net) outperformed all other methods in both evaluation metrics and computation time.

Conclusions:

• SDU-Net presents the best performance for 3D reconstruction from 2D grayscale images of laser-welded components.
• SDU-Nets require fewer trainable parameters and benefit from data augmentation, enabling robust models with minimal training data.
• Deep learning, particularly SDU-Nets, offers a promising approach for efficient and accurate 3D reconstruction in smart manufacturing applications.