Experiments in Fluids 24 (1998) 373—374 Springer-Verlag 1998
An artificial neural network for double exposure PIV image analysis
P.-H. Chen, J.-Y. Yen, J.-L. Chen
This note presents a back propagation neural
network for PIV image analysis. Unlike the conventional
auto-correlation method that identiﬁes one pair of image out of
the picture, the proposed network distinguishes all the image
pairs in the measurement area and provides different labels for
each pair. Experimental investigations show good agreement
with the auto-correlation process for the uniform ﬂow mea-
surement, and a 78.1% success ratio for the stagnation ﬂow.
This note investigates the use of a neural network for a double
exposure particle image velocimetry (PIV) ﬂow ﬁeld analysis.
The PIV system uses a two-dimensional auto-correlation
process to infer a proper pair from among the particle images
recorded on a photographic ﬁlm (Adrian 1991). Although the
auto-correlation process involves very intensive computation,
and it requires the assumption that the ﬂow has to be fairly
uniform (Anderson and Longmire 1996), little improvements
have been made. Most popular modiﬁcations seek to introduce
interpolation and extrapolation to widen its applications.
There is also an attempt to aid the analysis with the Particle
Tracking Velocimetry (PTV) technique, which is a redundant
practice limited by the equipment complexity (Anderson and
An alternative to the auto-correlation process to perform
image pairing is to use the artiﬁcial neural network (Grant and
Pan 1995; Carosone et al. 1995). The former results claim
60—97% success ratio depending on different degrees of ﬂow
turbulence and image densities. In this note, the authors try to
develop a more universal neural network for image pairing in
the PIV system. Unlike the former network which identiﬁes
one pair of particles in each ‘‘segment’’, the proposed network
takes all the input information simultaneously and attaches
separate labels to all the distinguishable image pairs. The
experimental results show that the proposed network achieves
a good success ratio, and is capable of velocity vectors in
P.-H. Chen, J.-Y. Yen and J.-L. Chen
Department of Mechanical Engineering, National Taiwan University,
Taipei, Taiwan 10764, R.O.C.
Correspondence to: J.-Y. Yen
relatively diverse directions. Thus, the uniform ﬂow ﬁeld
assumption for the two-dimensional auto-correlation process
is also somewhat relaxed.
The neural network for PIV image analysis
The PIV system shines two consecutive sheets of laser pulses
into the ﬂuid ﬂow and records the scattered particle images on
a photographic ﬁlm. As a result, each particle will leave two
images on the ﬁlm, and the distance between these images and
the relative positions determines the two-dimensional velocity
vector. The resolution for the 1 mm2 PIV interrogation area
is 512;512 pixels. For a good average velocity, this picture
is ﬁrst reduced to 32;32 pixels, and the neural network
processes a quarter picture, which are 16;16 pixels, at a time.
To improve the velocity estimation, the center of each particle
image is computed and recorded before the reduction.
To avoid ambiguity, the gray levels are thresh-held to 0 and
1, where a 1 would represent the appearance of a particle
center. The centers of the particles are calculated by averaging
the four edges of the images. Since the pictures are 512;512
pixels, the measurement resolution for distance is 2 m.
Fig. 1. The PIV image of the ﬂow in front of a moving turbine blade