Application of neural networks to stereoscopic imaging velocimetry

Stereoscopic imaging velocimetry is an optical nonintrusive method for meas uring three-dimensional three-component gross-field fluid flows that is bas ed on the images captured by two charge-coupled device sensors from differe nt vantage points. In this approach part of the individual particle images or equivalent data points are likely to be lost when a flowfield with a hig h-particle density is captured by the imaging system. The data loss and err oneous detection mostly occur during the process of overlap decomposition o f superimposed particle images and during the phase of particle tracking. T o maximize the data point recovery and to enhance the measurement reliabili ty by correctly identifying particles and tracks, neural networks are imple mented in the two phases of stereoscopic imaging velocimetry. For the phase of particle overlap decomposition, the back propagation neural network is used because of its ability in pattern recognition and nonlinear classifica tion. For the phase of particle tracking,the Hopfield neural network is emp loyed to attain a globally optimal solution in finding appropriate particle tracks. Our investigation indicates that the neural networks offer very go od potential for performance enhancement and has proven to be very useful f or stereoscopic imaging velocimetry.