Neural-network method applied to the stereo image correspondence problem in three-component particle image velocimetry

The successful application of a recurrent neural network of the Hopfield type to the solution of the stereo image-pair reconciliation problem in stereoscopic particle image velocimetry (PIV) in the tracking mode is described. The results of applying the network to both virtual-flow and physical-flow PIV data sets are presented, and the usefulness of this novel approach to PIV stereo image analysis is demonstrated. A partner-particle image-pair density (PPID) parameter is defined as the average number of potential particle image-pair candidates in the search window in the second view corresponding to a single image pair in the first view. A quantitative assessment of the performance of the method is then made from groups of 100 synthetic flow images at various values of the PPID. The successful pairing of complementary image points is shown to vary from 100% at a PPID of 1 and to remain greater than 97% successful for PPID's up to 5. The application of the method to a hydraulic flow is also described, with in-line stereo images presented, and the application of the neural-matching method is demonstrated for a typical data set.