Active particle feedback control with a single-shot detection convolutional neural network.

Area of Science:

• Microscopy
• Active Matter Physics
• Artificial Intelligence

Background:

• Real-time object detection in optical microscopy is crucial for manipulating microscopic entities.
• Conventional methods struggle with heterogeneous particle ensembles and low signal-to-noise ratios at video rates.
• Advanced detection techniques are needed for complex microscopic systems.

Purpose of the Study:

• To present a convolutional neural network (CNN) single-shot detector for real-time microscopic object detection.
• To enable localization and classification of multiple objects with varying contrast in video streams.
• To demonstrate the framework's adaptability for new particle types and parameters, such as orientation.

Main Methods:

• Development of a custom convolutional neural network single-shot detector.
• Application of the detector to optical microscopy images at video rates (up to 100 fps).
• Testing the network's performance on images up to 1024x1024 pixels, including low signal-to-noise scenarios.

Main Results:

• The CNN detector achieves real-time performance, localizing and classifying multiple microscopic objects.
• The system operates effectively at high frame rates (100 fps) and with large image sizes.
• Successful selective control of self-thermophoretic active particles within a heterogeneous ensemble was demonstrated.

Conclusions:

• The developed CNN detector is suitable for real-time applications in optical microscopy.
• The framework offers adaptability for diverse particle types and detection parameters.
• This approach facilitates new research into collective behavior in active matter through artificial interaction rules.