FlowTexNet: Fast Texture Synthesis for Massive Flow Field Visualization

Abstract

Flow field texture synthesis is a common and popular way to visualize flow fields. When massive flow fields are to be processed, existing algorithms based on line integral convolution (LIC) are not fast enough. In this paper, a new deep-learning-based method is proposed to synthesize flow textures for massive flow fields. Firstly, a deep neural network called FlowTexNet is built on the base of encoder-decoder architecture. Then the network is trained by flow textures generated by the original LIC algorithm. By this way, FlowTexNet can synthesize flow textures that have the same visualization effect as LIC textures. But FlowTexNet is much faster than the LIC algorithm. Test results show that the speedup of FlowTexNet is up to 450x when it is used to process massive flow fields and compared with the original LIC algorithm. Moreover, FlowTexNet can be applied to flow fields that are out of training, showing good generalization performance.