Experimental Analysis of U-Net and Mask R-CNN for Segmentation of Synthetic Liquid Spray

Abstract

In digital image processing, segmentation is a process by which we can partition an image based on some variables to extract necessary elements. Unlike typical objects, it is complicated to segment dynamic objects from a synthetic fluid dataset where properties like position and shape change over time. Experiments on image segmentation over this dataset are conducted using U-Net (semantic segmentation) and Mask R-CNN (instance segmentation) to compare their results. The training dataset is generated from seven labelled images through data augmentation. Training on 1000 and validating on 200 images, Mask R-CNN achieved more epochs quickly. Around 1000 epochs for Mask R-CNN and 500 epochs for U-Net, both models reached a similar result in terms of F1 score and can segment the object in the new images.