Application of Convolutional Neural Networks for the analysis of soap film surface

The quantitative characterization of thin-film dynamics is fundamental to understanding complex fluid-dynamic phenomena of a soap film. The chaotic nature, rapid evolution, and tendency of flow patterns on soap films to merge make their analysis based on image processing via traditional methods exceedingly difficult. This thesis presents a hybrid computational pipeline that integrates supervised and unsupervised deep learning with classic computer vision approaches to address this challenge. For semantic segmentation, a U-Net architecture was trained on a dataset of manually annotated masks, enhanced by a sophisticated localized data augmentation strategy and a weighted loss function. In parallel, a Convolutional Autoencoder was employed for unsupervised clustering via K-Means to identify different flow regimes. Finally, the pattern dynamics were quantified using an IoU-based tracking algorithm and a velocity field analysis with the Farnebäck method. The main results include the unsupervised identification of three distinct flow regimes. Furthermore, both the tracking and velocity field analyses outlined the soap film pattern density and surface velocity characteristics. In conclusion, this work proposes and validates a hybrid methodology that provides a robust tool for the quantitative and automated characterization of complex surface dynamics, demonstrating the effectiveness of integrating diverse techniques for the analysis of experimental data.