Defect 3D reconstruction with integrated bright-field and dark-field structured illumination microscopy based on...

Area of Science:

• Optical Engineering
• Metrology
• Computer Vision

Background:

• Quantitative 3D inspection of micro-to-nanoscale surface defects in optical components for high-energy laser systems presents significant challenges.
• Existing methods struggle with sparse point clouds and noise in imaging techniques like bright-field and dark-field structured illumination microscopy (BDSIM).

Purpose of the Study:

• To develop a novel inspection framework for high-precision 3D reconstruction of micro-to-nanoscale surface defects.
• To address limitations of low luminous flux, sparse data, and noise in BDSIM imaging.

Main Methods:

• Integration of bright-field and dark-field structured illumination microscopy (BDSIM) with deep learning-based 3D reconstruction.
• Development of the Att-PU-Net model, an enhanced point cloud upsampling network (PU-Net) incorporating self-attention and multi-scale feature fusion.
• Implementation of a hybrid training strategy with simulated and real-world defect data.
• Design of a composite loss function including chamfer distance, repulsion loss, and curvature consistency.

Main Results:

• Att-PU-Net demonstrated superior geometric accuracy and uniformity compared to marching cubes and contour filter algorithms (P2S: 0.5720 µm, NUC: 0.3230).
• Experimental validation showed a reconstruction accuracy of 0.6343 µm and a maximum depth error of 0.79 µm compared to white light interferometry (WLI).

Conclusions:

• The proposed BDSIM and Att-PU-Net framework provides an effective and reliable solution for high-precision 3D reconstruction of complex optical surface defects.
• The method overcomes limitations of traditional techniques, enabling accurate defect characterization in critical optical components.