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

Addressing the challenges in quantitative 3D inspection of micro-to-nanoscale surface defects in optical components for high-energy laser systems, this paper proposes a novel, to our knowledge, inspection framework integrating bright-field and dark-field structured illumination microscopy (BDSIM) with deep learning-based 3D reconstruction. To mitigate the limitations of sparse point clouds and inherent noise caused by the low luminous flux in BDSIM imaging, we developed the Att-PU-Net model, building upon the point cloud upsampling network (PU-Net) architecture. This model incorporates a self-attention mechanism to enhance the contextual perception of local geometric abruptness and employs a multi-scale feature fusion strategy to preserve fine topological details. To ensure robust generalization from simulation to reality, a hybrid training strategy combining procedurally generated and real-world defect samples is adopted. Furthermore, a composite loss function integrating chamfer distance, repulsion loss, and curvature consistency constraints was designed to significantly improve point distribution uniformity and edge sharpness. Simulations and comparisons between Att-PU-Net and the marching cubes, contour filter algorithms demonstrate that Att-PU-Net achieves an optimal balance between geometric accuracy and uniformity (P2S: 0.5720 µm, NUC: 0.3230). Experimental validation on actual optical damage reveals a reconstruction accuracy of 0.6343 µm and a maximum error of only 0.79 µm in defect depth compared with white light interferometry (WLI), confirming the method's effectiveness and reliability for high-precision 3D reconstruction of complex optical surface defects.