M3D: A Benchmark Dataset and Model for Microscopic 3D Shape Reconstruction.

Area of Science:

• Optics and Photonics
• Computer Vision
• Materials Science

Background:

• Microscopic 3D shape reconstruction using depth from focus (DFF) is vital for precision manufacturing, but current models struggle due to a lack of suitable datasets.
• Existing DFF datasets differ significantly from microscopic imaging in optical design, imaging principles, and scene characteristics, limiting model performance.

Purpose of the Study:

• To introduce a novel microscopic DFF dataset (M3D) and a corresponding deep learning network (M3DNet).
• To address the limitations of existing DFF datasets and models in microscopic 3D shape reconstruction tasks.

Main Methods:

• Constructed M3D dataset using a custom microscopic device, featuring 1,952 multi-focus image sequences with depth labels derived from the 3D TFT algorithm and validated against design values.
• Developed M3DNet, a frequency-aware, end-to-end network designed to handle challenges such as shallow depth-of-field and weak textures inherent in microscopic imaging.

Main Results:

• M3D dataset successfully bridges the gap between macroscopic and microscopic DFF applications.
• M3DNet demonstrates effective capture of rapid focus decay and improved performance on public DFF datasets through enhanced global feature extraction.
• M3DNet exhibits strong robustness in challenging, extreme imaging conditions.

Conclusions:

• The M3D dataset and M3DNet significantly advance the field of microscopic 3D shape reconstruction.
• These contributions enable wider applications of DFF technology in precision manufacturing and quality control.
• The developed resources are publicly available for further research and development.