Faster RCNN과 픽셀 분류 신경망을 이용한 다상유동에서의 기포 검출에 대한 연구 = Study on Bubble Detection in Multiphase Flow Using Faster RCNNs and Pixel Classification Networks

다국어 초록 (Multilingual Abstract)

Characterization of bubbles in gas-liquid multiphase flow is important to design and optimize industrial systems. To this end, multiple visual techniques, including image processing, have been developed to measure bubble size distributions. However, traditional image processing methods significantly influence experimental settings. Thus, an improved bubbled detection method based on convolutional neural networks (CNNs) is proposed herein. This method involves using a faster region-based CNN (Faster RCNN) to detect the bubble location and a pixel classification network to segment bubble shapes. Experiments were conducted using three types of bubble generators with various background intensities and volume flow rates. Furthermore, the backbone CNN architectures used in the Faster RCNN and the pixel classification network were compared. The model network is highly versatile, is capable of characterizing bubble information in real-time, and has the potential to replace the traditional image processing method.

국문 초록 (Abstract)

다상유동에서 기포 크기 분포를 알기 위해 사용되고 있는 기존 이미지 처리 기법들은 실험 수행 조건에 영향을 받는다. 본 연구에서는 CNN을 기반으로 한 기포 검출 기법을 제안한다. 이미지에서 기포가 위치한 영역을 찾기 위해서 Faster RCNN을 사용하였고, 픽셀 분류 신경망을 이용해서 찾아진 영역에서 기포를 분할하였다. 신경망들을 훈련시키기 위해 필요한 이미지는 세 종류의 기포발생기를 이용한 실험으로부터 얻었으며, 배경 밝기, 유량들을 변경하며 실험하였다. 신경망의 backbone 구조에 따른 성능에 대한 비교를 진행하였다. 신경망의 깊이에 따른 구조 병렬화 영향에 대한 분석할 수 있었고 제안된 신경망은 실시간으로 이미지로부터 기포를 검출할 수 있었다.

참고문헌 (Reference)

1 Grau, R. A., "Visual Technique for Measuring Bubble Size in Flotation Machines" 15 (15): 507-513, 2002

2 Simonyan, K., "Very Deep Convolutional Networks for Large-scale Image Recognition"

3 Girshick, R., "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation" 580-587, 2014

4 Xie, C. G., "Measurement of Multiphase Flow Water Fraction and Water-Cut" 914 (914): 232-239, 2007

5 Goodfellow, I. J., "Generative Adversarial Nets" 2672-2680, 2014

6 Ren, S., "Faster RCNN: Towards Real-Time Object Detection with Region Proposal Networks" 91-99, 2015

7 Hampel, U., "Experimental Ultra Fast X-ray Computed Tomography with a Linearly Scanned Electron Beam Source" 16 (16): 65-72, 2005

8 Dickin, F., "Electrical Resistance Tomography for Process Applications" 7 (7): 247-, 1996

9 He, K., "Deep Residual Learning for Image Recognition" 770-778, 2016

10 Ilonen, J., "Comparison of Bubble Detectors and Size Distribution Estimators" 101 : 60-66, 2018

11 Schäfer, R., "Bubble Size Distributions in a Bubble Column Reactor under Industrial Conditions" 26 (26): 595-604, 2002

12 Fu, Y., "BubGAN : Bubble Generative Adversarial Networks for Synthesizing Realistic Bubbly Flow Images" 204 : 35-47, 2019

13 Haas, T., "BubCNN: Bubble Detection Using Faster RCNN and Shape Regression Network" 216 : 115467-, 2020

14 Pelikan, M., "BOA: The Bayesian Optimization Algorithm" 1 : 525-532, 1999

15 Poletaev, I. E., "Artificial Neural Network for Bubbles Pattern Recognition on the Images" 754 (754): 072002-, 2016

16 Xie, S., "Aggregated Residual Transformations for Deep Neural Networks" 1492-1500, 2017

17 Prasser, H. M., "A New Electrode-Mesh Tomograph for Gas-Liquid Flows" 9 (9): 111-119, 1998