Solving Level Set Evolving Using Fully Convolution Network

We have designed and trained a fully convolution neural network to calculate geometry evolving in discontinuous velocity fields. Train data set is generated by numerically solving level set equations. The equation-based regression problem is innovatively transformed to a series of classification problem to make the training easier. A network structure similar to deep residual net is used to improve accuracy. Test shows that our network is 18x faster compared to standard level set method, and error bound of our work is 0.5% for single frame prediction, and 9% for iterative prediction. Accuracy may be further improved by using priori knowledge or introducing generative adversarial network.