BubCNN: Bubble detection using Faster RCNN and shape regression network. (28th April 2020)

• Record Type: Journal Article
• Authors: Haas, Tim; Schubert, Christian; Eickhoff, Moritz; Pfeifer, Herbert
• Is Part Of: Chemical engineering science. Volume 216(2020)
• Publication Date: 2020-04-28
• Subjects: Bubble detection -- Machine learning -- Convolutional neural networks -- Image processing; Chemical engineering -- Periodicals; Génie chimique -- Périodiques; Chemical engineering; Periodicals; Electronic journals; 660
• DOI: 10.1016/j.ces.2019.115467
• Languages: English
• ISSNs: 0009-2509

Abstract

Graphical abstract: Highlights: Development of a bubble detection method based on machine learning named BubCNN. Includes a Faster RCNN detector module and a regression CNN for shape approximation. Optimal Hyperparameters and architecture for both modules were analyzed. BubCNN yield accurate results and show great potential for generalization. Detector and a Transfer Learning Module are made publicly available on GitHub. Abstract: Detailed knowledge about gas-liquid multiphase flows is important to optimize industrial systems. Imaging with image processing is the most commonly used measurement technique. However, the workflow and parameters strongly depend on the experimental conditions and no generally applicable process has been developed yet. Here, a workflow based on convolutional neural networks (CNN) is proposed that can be used with a wider range of experimental conditions. The method, named BubCNN, employs a Faster region-based CNN (RCNN) detector to locate bubbles and a shape regression CNN to predict bubble shape parameters. Hyperparameters and network architectures for both modules were systematically analyzed. BubCNN achieved accurate results for different experimental conditions. A pretrained program was made publicly available on GitHub. Since the whole variety of bubble images was not yet captured in the training data set, an additional semi-automatic transfer learning module is provided that allows to customize BubCNN for different images.