Related Experiment Video
Updated: May 8, 2025
15:04
Picometer-Precision Atomic Position Tracking through Electron Microscopy
Published on: July 3, 2021
6.3K
Precision autofocus in optical microscopy with liquid lenses controlled by deep reinforcement learning.
Jing Zhang1, Yong-Feng Fu2, Hao Shen1
1School of Mechanical and Electrical Engineering, Soochow University, No.8 Jixue Road, Suzhou City, Jiangsu, 215000, China.
Microsystems & Nanoengineering
|December 24, 2024
Summary
This study introduces an adaptive liquid lens microscope system using Deep Reinforcement Learning-based Autofocus (DRLAF) for faster, more precise microscopic imaging. The novel DRLAF approach significantly enhances autofocus speed and generalization capabilities for diverse samples.
Area of Science:
- Microscopy and Imaging Technologies
- Artificial Intelligence in Scientific Instrumentation
- Optical Engineering
Background:
- Microscopic imaging is vital across science and engineering, but current systems face limitations in miniaturization and autofocus speed.
- Traditional autofocus methods struggle with hardware constraints and slow software processing, hindering rapid, precise imaging.
- There is a critical need for advanced autofocus techniques to improve the efficiency and applicability of microscopic systems.
Purpose of the Study:
- To develop and evaluate an adaptive liquid lens microscope system integrated with Deep Reinforcement Learning-based Autofocus (DRLAF).
- To achieve end-to-end autofocus by training a deep reinforcement learning agent directly on image data.
- To enhance autofocus speed, robustness, and generalization capabilities compared to existing methods.
Main Methods:
- Implementation of a custom liquid lens as an 'agent' in a reinforcement learning framework, with images as 'states' and voltage adjustments as 'actions'.
- Development of a targeted reward function to optimize autofocus performance, moving beyond simple sharpness assessment.
- Utilizing parallel 'state' dataset lists with random sampling for enhanced model adaptability and generalization to unknown samples.
Main Results:
- The Deep Reinforcement Learning-based Autofocus (DRLAF) system achieved an average autofocus speed of 3.15 time steps.
- Demonstrated a significant 79% increase in speed compared to traditional search algorithms.
- Attained a high success rate of 97.2% and superior generalization capabilities over other deep learning methods.
Conclusions:
- The proposed liquid lens microscope system with DRLAF offers a robust and efficient solution for rapid, precision autofocus.
- DRLAF significantly improves autofocus speed and accuracy, addressing key limitations of traditional microscopic imaging systems.
- The enhanced generalization capability makes the system highly adaptable for analyzing a wide range of unknown samples.
