Flow structure segmentation for vortex identification using butterfly convolutional neural networks

Vortex identification is important for understanding the physical mechanism of turbulent flow. The common vortex identification techniques based on velocity gradient tensor such as Q criterion will consume a lot of computing resources for processing great quantity of experimental data. To improve the vortex identification efficiency and achieve real-time recognition, we present a novel vortex identification method using segmentation with convolutional neural network (CNN) based on flow field image data, which is named "Butterfly-CNN". Considering that the view of flow field is small, it is necessary to integrate both the local and global feature maps to achieve higher precision. The architecture consists of an encoded-decoded path, which is similar to U-net but with different superimposed network part. In the Butterfly-CNN, the cross-expanding paths are designed with the global information to enable precise localization, and the feature maps after each convolution are regarded as the original pictures, then convolute to the size of the last feature map and upsample to the original size again. Finally, the decoded and cross-expanding networks are added up. The Butterfly-CNN can be trained end-to-end from a few images, and it is useful and efficient for vortex identification.