A comparison of YOLOv8 and Mask R-CNN and subsequent object tracking algorithms to identify and reconstruct bubbles in two-phase flow.
Computer Vision, Wire Mesh Sensor (WMS), two-phase gas liquid flows, machine learning, instance segmentation, YOLOv8, Detectron2, Mask R-CNN, object tracking
The primary goal of this project was to leverage deep learning methodologies to accurately identify and reconstruct gas bubbles in two-phase gas-liquid flows. Specifically, the project aimed to utilize data captured from a Wire Mesh Sensor (WMS) to individually identify and segment bubbles, showcasing the efficacy of deep learning and object tracking algorithms in processing WMS data for enhanced analysis and understanding of gas-liquid interactions.
The current process of reconstructing bubbles involves multiple sequential steps that required external inputs and manual analysis, which is time-consuming and potentially prone to errors. Two deep learning architectures, YOLOv8 and Mask R-CNN, were selected and compared to tackle the challenge of accurately detecting bubbles of different sizes and shapes, using data converted into images from the WMS.
The application of deep learning methods yielded promising results, with the identification of up to 88% of the gas contained within the liquid using the YOLO architecture. Furthermore, implementing simple object tracking algorithms, like centroid tracking, demonstrated satisfactory outcomes in bubble reconstruction. These achievements highlight the potential of combining deep learning models with object tracking algorithms to streamline the process of bubble reconstruction from raw WMS data, ultimately facilitating a more efficient and accurate analysis of gas-liquid flows.
September 23 - March 24, 360h p.p., Team of 2
Paul Scherrer Institute (PSI), Forschungsstrasse 111, 5232 Villigen PSI
Navjot Zubler, Lukas Reber
Prof. Dr. Arzu Cöltekin (arzu.coltekin@fhnw.ch), Dr. Leticia Fernà ndez Moguel (leticia.fernandezmoguel@fhnw.ch)
