LIF based detection of low-speed streaks
A. Agrawal, L. Djenidi, R. A. Antonia
Abstract An automated method for detecting low-speed
streaks in laser-induced ﬂuorescence (LIF) images of tur-
bulent boundary layers is described. The method, based on
the identiﬁcation and characterization of local maxima in
intensity educes all the low-speed streaks that can be
visually identiﬁed in the LIF images. The proposed algo-
rithm seems robust and can yield reliable information on
the statistical characteristics of the streaks.
Low-speed streaks are a characteristic feature of the near-
wall region of a turbulent boundary layer ﬂow (see e.g.
Corrsin 1957; Kline and Robinson 1988). Several studies
have been undertaken in the past to document the length
and spacing of these streaks (e.g. Smith and Metzler 1983;
Klewicki et al. 1995). However, the number of streaks that
can be examined in these studies is rather small because of
the nature of the techniques that were adopted for locating
the streaks. For example, Smith and Metzler marked the
location of streaks on a transparency, while Klewicki et al.
used a ruler on the video monitor displaying the images.
Clearly, both methods were tedious and not free of sub-
jectivity. It is possible that the large scatter in the values of
the streak spacing and length (50–300 and 600–1000 wall
units respectively; Smith and Walker 1997) reﬂected the
inaccuracies in the detection procedure and the small size
of the data set. These concerns are sufﬁcient motivation
for developing a robust technique for detecting low-speed
streaks. An automated method, which is accurate, more
convenient and less subjective than the previous methods,
is presented here.
We are not aware of any previous attempts to develop
and automate a streak detection procedure. Streaks occur
intermittently in space and time, and they do not have
clear edges. Therefore detection of streaks is difﬁcult
through standard image processing software. Djenidi et al.
(2002) proposed the idea of automated streak detection to
investigate the effect of concentrated wall suction on the
low-speed streaks. However, the procedure involved too
many adjustable constants, making it somewhat sub-
jective. The present detection method, which contains only
one adjustable constant, represents an improvement over
the procedure outlined in Djenidi et al. Analysis of a large
dataset is possible using this procedure, thereby providing
useful statistics for several quantities of interest.
Detection algorithm and results
The algorithm determines ﬁrst the centre of a streak, and
then computes its width and length. For most of the steps
outlined below, the images are scanned ﬁrst horizontally
(along the spanwise direction) and then vertically
(streamwise direction). The centre of a streak is deter-
mined with the following procedure:
1. The noise in the raw image is removed (low-pass
ﬁltered) using wavelets.
2. The intensity gradient along a row is computed after a
small amount of smoothing (using a moving-point
average) of the image. This additional step of ﬁltering
is proposed because some LIF images can be of
modest quality. It was veriﬁed that this step does not
suppress the signal peaks. Instead, it only removes the
remaining high-frequency jitter before differentiating.
3. Locations of zero gradient are identiﬁed. If they also
correspond to large local values in the ﬁltered image,
i.e. greater then the mean of the values in the vicinity
of the point considered, then the point is accepted as a
peak. This latter condition is required to avoid
detecting small side peaks associated with the primary
peak, and very closely spaced peaks. Note that a peak
can be deleted if certain later conditions (at step 11)
are not satisﬁed.
The above steps have been illustrated in Fig. 1, which
shows a raw intensity proﬁle along a given row. The ﬁl-
tering process (step 1) removes most of the high-frequency
noise in the original proﬁle (the proﬁles have been dis-
placed upwards for clarity). This step allows the detection
of small peaks that are hardly discerned in the original
proﬁle. However, because of the remaining noise, the ﬁl-
tered proﬁle cannot be differentiated, and further
smoothing is needed (step 2). It is reassuring that
smoothing does not affect the peaks in any undesirable
Experiments in Fluids 36 (2004) 600–603
Received: 5 June 2003 / Accepted: 7 October 2003
Published online: 3 March 2004
Ó Springer-Verlag 2004
A. Agrawal (&), L. Djenidi, R. A. Antonia
Department of Mechanical Engineering, University of Newcastle,
2308, Callaghan, NSW, Australia
We are grateful to P.-E. Gall and A. Vincent for providing the LIF
data used in the paper. RAA and LD acknowledge the support of
the Australian Research Council.