Periodic Reporting for period 1 - DROPTRACK (Fast & automated droplet tracking tool for microfluidics)
Reporting period: 2022-10-01 to 2024-03-31
Microfluidic applications require investigation of droplet dynamics and morphologies. This is particularly important for processes like drug delivery and production of cosmetics and paints, where droplet shape evolution is crucial. Traditionally, specialized equipment was needed to quantify droplet size, packing fraction, flow rates etc. We present DropTrack, a computer vision tool that infers these physical parameters by analyzing droplet images from experiments and production lines. DropTrack can detect droplets in static images and track their motion in videos. The extracted data allows for the calculation of droplet morphologies and key physical parameters like their sizes and packing fraction, which are essential for quality control in microfluidic manufacturing.
DropTrack utilizes the YOLO (You Only Look Once) and DeepSORT algorithms for droplet identification and tracking. We trained these algorithms using a dataset comprised of approximately 2,000 manually annotated images obtained from microfluidic experiments, supplemented with another 2,000 synthetically generated images. These images contained labelled droplets. The dataset was then employed to train multiple YOLOv5 and YOLOv7 models alongside the DeepSORT algorithm. The resulting trained networks achieve robust droplet detection under diverse conditions and generate their trajectories. This data enables the computation of crucial physical parameters such as packing fraction, droplet size distribution, flow rate, and packing order. Both the training dataset and the trained networks are freely available for further development.
The project emphasizes two primary outcomes. Firstly, a comprehensive dataset comprising droplets and corresponding annotations is provided, facilitating the training of various computer vision models. Dataset compilation represents a foundational task in the development of computer vision applications. To the best of our knowledge, there exists no other dataset containing labeled droplet images specific to microfluidics, apart from the one curated within this project. Secondly, we release pre-trained weights for YOLOv5, YOLOv7, and DeepSORT, all trained on this dataset. These weights enable the tracking of droplets within a given sequence of images across a spectrum of microfluidics manufacturing processes and experimental setups.
In subsequent iterations, DropTrack could be enhanced to incorporate additional functionalities, such as droplet boundary identification (masking) and the mapping of droplet shapes to minimal representations. These enhancements are crucial for enabling self-learning algorithms to autonomously control microfluidics production processes. These functionalities will be pivotal for the development of commercial software prior to market entry.
In subsequent iterations, DropTrack could be enhanced to incorporate additional functionalities, such as droplet boundary identification (masking) and the mapping of droplet shapes to minimal representations. These enhancements are crucial for enabling self-learning algorithms to autonomously control microfluidics production processes. These functionalities will be pivotal for the development of commercial software prior to market entry.