Fast autofocusing using tiny transformer networks for digital holographic microscopy.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Digital holography offers extended focus without mechanical z-axis movement.
• Accurate determination of focal distance in digital holography is challenging and time-consuming.
• Deep learning (DL) presents a potential solution for automated focusing.

Purpose of the Study:

• To develop and evaluate a DL-based autofocusing method for digital holography.
• To investigate the performance of tiny DL models for focal distance regression.
• To compare the proposed tiny models against established neural networks.

Main Methods:

• Digital holograms were acquired using a Digital Holographic Microscope (DHM) with a 10x objective.
• Single wavelength holograms of a patterned target over a 92 μm axial range were used.
• Tiny Vision Transformer (TViT), tiny VGG16 (TVGG), and tiny Swin-Transformer (TSwinT) models were proposed and tested.

Main Results:

• Tiny DL models accurately predicted focusing distance (ZRPred) with an average accuracy of 1.2 μm experimentally.
• Numerical simulations showed errors below 0.3 μm for all tiny models.
• All models achieved inference times under 25 ms on CPU, with TViT showing robustness to occlusions.

Conclusions:

• DL-based autofocusing offers a significant improvement for digital holography.
• Tiny DL models provide an efficient and accurate solution for focal distance determination.
• This approach enhances computer vision capabilities in 3D microscopy and micro-robotics.