BYG-drop: a tool for enhanced droplet detection in liquid–liquid systems through machine learning and synthetic imaging

A new image processing machine learning algorithm for droplet detection in liquid–liquid systems is here introduced. The method combines three key numerical tools—YOLOv5 for object detection, Blender for synthetic image generation, and CycleGAN for image texturing—and was named “BYG-Drop for Blender...

Full description

Saved in:

Bibliographic Details
Published in:	Frontiers in chemical engineering Vol. 6
Main Authors:	Bana, Grégory, Lamadie, Fabrice, Charton, Sophie, Randriamanantena, Tojonirina, Lucor, Didier, Sheibat-Othman, Nida
Format:	Journal Article
Language:	English
Published:	Frontiers Media 08-08-2024 Frontiers Media S.A
Subjects:	convolutional neural networks (CNNs) droplet detection droplet size distribution Engineering Sciences generative adversarial networks (GANs) liquid-liquid emulsion machine learning convolutional neural network CNN generative adversarial network GAN machine learning droplet size distribution droplet detection
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	A new image processing machine learning algorithm for droplet detection in liquid–liquid systems is here introduced. The method combines three key numerical tools—YOLOv5 for object detection, Blender for synthetic image generation, and CycleGAN for image texturing—and was named “BYG-Drop for Blender-YOLO-CycleGAn” droplet detection. BYG-Drop outperforms traditional image processing techniques in both accuracy and number of droplets detected in digital test cases. When applied to experimental images, it remains consistent with established techniques such as laser diffraction while outperforming other image processing techniques in droplet detection accuracy. The use of synthetic images for training also provides advantages such as training on a large labeled dataset, which prevents false detections. CycleGAN’s texturing also allows quick adaptation to different fluid systems, increasing the versatility of image processing in drop size distribution measurement. Finally, the processing time per image is significantly faster with this approach.
ISSN:	2673-2718 2673-2718
DOI:	10.3389/fceng.2024.1415453