High-precision microscopic autofocus with a single natural image.

Area of Science:

• Microscopy
• Image Processing
• Machine Learning

Background:

• Learning-based autofocus methods are crucial for rapid, high-quality image acquisition in industrial microscopy.
• Existing methods suffer from fitting and dataset errors, limiting focusing accuracy.
• Improving autofocus precision is essential for advanced microscopic analysis.

Purpose of the Study:

• To introduce a high-precision autofocus pipeline for predicting defocus distance from single natural images.
• To develop a novel dataset generation method that overcomes sharpness metric limitations.
• To enhance focusing accuracy using a lightweight regression network.

Main Methods:

• A new dataset creation methodology was implemented, addressing limitations of traditional sharpness metrics.
• A lightweight regression network, the Natural-image Defocus Prediction Model (NDPM), was developed.
• A realistic, large-scale dataset was generated to train the models.

Main Results:

• The proposed dataset generation method improved overall dataset accuracy.
• The NDPM demonstrated superior focusing performance compared to other models.
• The mean focusing error achieved by NDPM was 0.422µm.

Conclusions:

• The developed autofocus pipeline and NDPM significantly enhance focusing accuracy in industrial microscopy.
• The novel dataset creation method provides a more robust foundation for training autofocus models.
• This work offers a promising solution for precise microscopic imaging applications.