3D Wire Reconstruction from Focal Stack Images Using Deep Learning

Rekoert, Sam

3D Wire Reconstruction from Focal Stack Images Using Deep Learning

Files

Rekoert, S_s1010799_MSc_MKI94_Thesis_2025.pdf (16.03 MB)

Authors

Rekoert, Sam

Issue Date

2025-02

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/20711

Abstract

Abstract This thesis addresses the challenge of reconstructing three-dimensional wire structures from focal stack images in semiconductor assembly inspection. Wire bonding, a critical process in semiconductor packaging, requires precise quality control that depends on accurate three-dimensional visualization and analysis. While focal stack imaging provides rich depth information, converting these multi-layered images into accurate 3D representations remains challenging. We present an approach to this problem, developing and evaluating various 3D convolutional neural network architectures adapted from medical image analysis techniques. Our methodology consists of the development of a synthetic data generation pipeline using Blender, creation of GPU-accelerated voxelization processes for ground truth generation, and implementation of multiple neural network architectures including variants of 3D U-Net with attention mechanisms and deep supervision. We evaluate these architectures using a dataset of 866 unique wire configurations, each comprising 32-plane focal stacks with dimensions of 512×512×32. Results demonstrate that while more complex architectures achieve higher accuracy (Dice score of 0.8068 for SEUNet with deep supervision), lighter architectures like HalfUNet achieve comparable performance (0.7953) with only 10% of the parameters. Our analysis reveals strong performance in x-y plane reconstruction but identifies challenges in depth estimation, particularly in real-world applications where image quality deteriorates with depth. The study also highlights the impact of different loss functions and feature scale configurations on reconstruction quality and computational efficiency. These findings provide valuable insights for industrial applications, identifying both the potential and limitations of deep learning approaches in wire bond inspection. While the results demonstrate promising capabilities in synthetic environments, they also reveal important considerations for bridging the gap between simulated and real-world performance, suggesting directions for future research in industrial inspection systems.