A machine learning approach to segment images of foam at a transparent wall.

This article describes the use of a machine learning based technique
to measure the bubble sizes of foam with small liquid fraction in contact with a
transparent wall. For two different experimental cases images are obtained of foam
in a cylindrical column and labeled with a classical image processing algorithm. An
available neural network based model, initially designed for cell image applications,
is trained and validated to segment the images. When comparing the bubble size
distribution in images found using the trained model with manually segmented images
a good agreement over a large range of diameters can be found. The error of the mean
diameter in both cases lies below 10%, mostly attributed to the failed recognition of
tiny round bubbles in dry foam. The trained model is provided for further usage.