Colour class identification of tracers using artificial neural networks

In this presentation a multilayer perceptron is used to classify coloured tracers. In fluid mechanics a nonintrusive measuring method delivering experimental information with a Lagrangian point of view (i.e. following the flow) would be extremely useful to clarify the origin, birth and development of vortical structures in technical systems. For this purpose Particle-Tracking-Velocimetry (PTV) might be employed. In PTV small tracers are tracked by a multi camera setup over time. With the known position of the tracers in at least two camera images it is possible to compute the 3d position of a tracer in space. In doing so it is difficult to solve the temporal and the spatial correspondence problem at high tracer density. Using coloured tracer particles the problem becomes much easier because the colour information can be used to support the correspondence analysis. To recognise the colour of particles, single chip cameras with a Bayer-Pattern are used. Because of the small diameter of the employed tracers (<0.1 mm), conventional interpolation methods do not work to reconstruct the colour information. Therefore, a multilayer perceptron with one or more hidden layers is employed to assign the tracers to their colour class. The feature vector of a tracer consists of the raw black/white-data of the Bayer-sensor as well as of structural attributes, such as the position of the tracer in relation to the camera pixel elements. The feature vector contains finally about 10 elements. In our example we have 4 colour classes. A training data set for one class has about 8000 feature vectors. The backpropagation-training converges in about 250 steps. The computational time of the recall is negligible. After training, the network is able to assign correctly about 90% of the tracers in each colour class.