An effective convolutional neural network for liquid phase extraction in two-phase flow PIV experiment of an object entering water

In the two-phase flow particle image velocimetry (PIV) experiment of an object entering water, the accurate extraction of the liquid phase region is an important step. In this paper, we elaborately design an effective convolutional neural network (CNN) called LTPNet to solve the problem of two-phase-flow boundary segmentation. Considering the supervised learning strategy, we make a dataset which is based on the two-phase flow PIV experiment of an object entering the water. The experimental results show that LTPNet can achieve high segmentation precision (above 0.98 DSC) on the test images. Meanwhile, our approach can have high computational efficiency with only 2.61M parameters and a speed of 17.34 ms on a single GTX 1080Ti card. Furthermore, the cross-correlation algorithm WIDIM is used to process the images segmented by LTPNet. The results show that our method can efficiently achieve the segmentation of two-phase-flow images and reduce the error vector of the phase edge.