APPLICATION OF NEURAL NETWORKS TO QUANTITATIVE FLOW VISUALIZATION
Ichirô KIMURA, Yasuaki Kuroe, Mamoru Ozawa · Journal of Flow Visualization and Image Processing · 1993
It is an important challenge to analyze a highly complex flow field in engineering, science, agriculture, and medicine. For such an analysis, it is essential to measure instantaneous physical quantities over the entire flow field. Recently, some flow measurement systems have been developed using image processing to obtain useful information from visualized flow images. The technique, however, has some problems. This report presents two new algorithms using neural networks for improving the measured values obtained by the image processing. One is a new algorithm for formularizing the relationship between color and temperature using the Back-propagation neural networks. The formularization is needed for the analysis of thermal flows using a thermo-sensitive liquid crystal method. The algorithm extends the measurement range. The other new algorithm uses the Hopfield neural networks for determining erroneous vectors, which appear in a velocity vector distribution estimated by a correlation method. The Hopfield network can distinguish erroneous from correct vectors even if the density of the erroneous vectors in a flow field is high.