The Generation of Labelled Bubbly Flow Images Using Cycle-Consistent Adversarial Network

The recent application of machine learning methods has provided new insights for particle/bubble detection and segmentation in fluid mechanics research. However, the training of artificial neural networks usually requires a formidable amount of high-quality labeled data which are tedious to acquire. In this work, we propose an unsupervised approach based on Cycle-Consistent Generative Adversarial Networks (CycleGAN to generate realistic bubbly flow images from given mask images. Namely, a mapping based on a group of unpaired bubble images of interest and masks is trained to translate a given mask to a bubbly flow image with the image closely following the geometric information of the mask. Compared to previous GAN-based methods, the proposed architecture can generate more realistic bubbly flow images by considering the differences in bubble images due to the change in the experimental setting. We test the accuracy of the method using both BubGAN data and labelled experimental data. For both cases, realistic bubbly flow images closely following the geometric properties of mask images are generated. The root-mean-square geometric error for single bubble generation using BubGAN data is about 0.85%. Our method can greatly reduce the cost of labeling images and the generated images can be used as benchmark or training data for other experimental data processing algorithms. Moreover, the method is not limited to bubbly flows and can be easily adapted to other particle-laden flow applications.