Overlapping bubble detection and tracking method based on convolutional Neural network and Kalman Filter. (14th December 2022)

• Record Type: Journal Article
• Authors: Wen, Daizhou; Chen, Wuguang; Yin, Junlian; Song, Yuchen; Ren, Mingjun; Wang, Dezhong
• Is Part Of: Chemical engineering science. Volume 263(2022)
• Publication Date: 2022-12-14
• Subjects: Overlapping Bubbles -- Image Processing -- Convolutional Neural Networks (CNN) -- Kalman Filter (KF); Chemical engineering -- Periodicals; Génie chimique -- Périodiques; Chemical engineering; Periodicals; Electronic journals; 660
• DOI: 10.1016/j.ces.2022.118059
• Languages: English
• ISSNs: 0009-2509

Abstract

Highlights: An identification method of overlapping bubbles in high void fraction conditions by Convolutional Neural Network was proposed. A trajectory tracking technology for overlapping bubbles was achieved based on Kalman Filter. The model was trained only by synthetic images generated by GAN. Overlapping bubbles can be detected and tracked accurately under low illumination and strong noise conditions. Abstract: Gas-liquid bubbly flow is widely applied in chemical process engineering. Geometric and dynamic parameters of bubbles play an essential role in the numerical prediction of mass and heat transfer processes. However, the critical obstacle in bubble detection is the inability of bubble segmentation and reconstruction when the overlapping issue of multiple bubbles is serious under high void fraction conditions. A new detection and tracking technique for overlapping bubbles was proposed in this paper to identify the overlapped bubbles. First, a novel convolutional neural network is used to detect bubbles. Afterward, the relationship between the detected bubbles in two frames is correlated using the Kalman Filter and neural network. The algorithm achieves 85 % accuracy under high overlap rate conditions in a 10 mm narrow rectangular channel with around 0.1 s for an image. In addition, a comparison test was conducted to evaluate the present technique's accuracy and robustness compared with conventional methods.