Development of a deep learning-based image processing technique for bubble pattern recognition and shape reconstruction in dense bubbly flows

This work presents a Convolutional Neural Network (CNN) based method for the shape reconstruction of bubbles in bubbly flows using high-speed camera images. The bubble identification and shape reconstruction adopted a methodology based on a set of anchor points and boxes, where a single anchor point is used for different anchor boxes with various sizes. These anchor points are determined, based on the internal features of the bubble images, which are more easily identifiable, in particular, in regions of the images with high bubble overlapping. This makes possible the application of the procedure to high void fraction bubbly flows. For a given anchor point, different ellipsoidal shapes are suggested as bubble shape candidates and are then correctly chosen by a trained CNN. The CNN training used labeled images from air–water system data set and a hyper-parameter analysis was performed to find the best configuration of the CNN architecture. From this optimal CNN architecture, different high-speed camera acquisitions of bubbly flows were analyzed by the CNN-based bubble shape reconstruction method. In order to gain a better comprehension of the method, experiments were conducted in two gas–liquid systems, air–water and air-aqueous glycerol solution, which resulted in different image parameters, such as brightness, contrast and edge definition. The CNN method trained only with air–water data, showed excellent performance in the cases with air-aqueous glycerol, demonstrating its generalization capability. In addition, the results showed that the deep learning method used in this work is able to detect most of the bubbles present in the high-speed camera images, even in dense bubbly flow configurations. The method developed in this work can be used to further analyze bubbly flows and generate experimental data for the implementation and validation of CFD models.