Multiresolution classification of turbulence features in image data through machine learning

• Published in: Computers & fluids Vol. 214; p. 104770
• Main Authors: Pulido, Jesus, da Silva, Ricardo Dutra, Livescu, Daniel, Hamann, Bernd
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 15.01.2021; Elsevier BV; Elsevier
• Subjects: Cost analysis; Data analysis; Data reduction; Fluid dynamics; Fluid flow; Histograms; Image classification; Image databases; Image filters; Image processing; Machine learning; MATHEMATICS AND COMPUTING; Support vector machines; Turbulence; Vortex detection; Vortices; Turbulence; Image processing; Vortex detection; Machine learning
• ISSN: 0045-7930; 1879-0747
• DOI: 10.1016/j.compfluid.2020.104770

•Intermediate data products for large simulations allow for meaningful analysis.•An image-space descriptor for the detection and extraction of vortex-like features.•A complete training and classification system enables low-cost evaluation of data.•Vortex features are detected at multiple sizes and scales through machine learning. During large-scale simulations, intermediate data products such as image databases have become popular due to their low relative storage cost and fast in-situ analysis. Serving as a form of data reduction, these image databases have become more acceptable to perform data analysis on. We present an image-space detection and classification system for extracting vortices at multiple scales through wavelet-based filtering. A custom image-space descriptor is used to encode a large variety of vortex-types and a machine learning system is trained for fast classification of vortex regions. By combining a radial-based histogram descriptor, a bag of visual words feature descriptor, and a support vector machine, our results show that we are able to detect and classify vortex features at various sizes at multiple scales. Once trained, our framework enables the fast extraction of vortices on new, unknown image datasets for flow analysis.