![](bg2.png)
its trajectory over time, PTV (Particle Tracking Ve-
locimetry) is used (Scharnowski and K
¨
ahler, 2020;
Ohmi and Li, 2000). Usually, a classical PTV al-
gorithm applied on pre-processed images consists of,
first, detecting the positions of each individual par-
ticle and then, matching and tracking these particles
across the image frames (Baek and Lee, 1996; Kim
and Lee, 2002). Over the years, different edge de-
tectors have been used to identify particles. Initially,
the particles were identified using single and dynamic
threshold binarization methods. Going forward, the
two-dimensional Gaussian regression technique was
applied to the particle intensity values to estimate the
sub-pixel particle centroid positions (Ohmi and Li,
2000; Heyman, 2019). However, with the increase in
the complexity in the images, when the intensity and
size of the particles varied, researchers went ahead
to apply different detectors based on noise, gradi-
ent, template, and morphology. These edge detectors
like Sobel, Canny, Robert, Prewitt, Gaussian, Over-
lap, and so on, are generally sensitive to the change
in pixel gray levels (Katiyar and Arun, 2014). Some
of these detectors output different results in terms of
their sensitivity towards noise and in detecting false
edges. Depending on the type of data input, differ-
ent applications of these detectors might work better.
Some will be good for larger and intense particles,
whereas others will have a better chance of catching
smaller and lighter ones (Janke et al., 2020). In or-
der to track particles across image frames acquired
over a time period, most of the algorithms are in-
fluenced by the probability relaxation algorithm tak-
ing into account the similar displacements exhibited
by the nearest neighboring particles (Baek and Lee,
1996; Ohmi and Li, 2000). The particle matching and
tracking were improved by using iterative matching
schemes and Deep Learning networks (Janke et al.,
2020; Heyman, 2019; Lee et al., 2019). However,
with these probabilistic methods, it is very difficult to
track multiple particles due to the occlusion/overlap
of two or more particles (Qian et al., 2021). In this
paper, we seek to improve the detection of particles
for PTV applications. For this purpose, different de-
tection methods are implemented, using Laplacian of
Gaussian (LOG) and Difference of Gaussian (DOG),
and compared to the traditional threshold binarization
method (Lefta et al., 2022). Noise minimization has
also been implemented, and finally, we also seek to
solve as much as possible the problem of overlapping
particles visualized during the detection and tracking
of the latter.
2 METHODOLOGY
In this paper, we have proposed an algorithm to de-
tect featureless micro or nanoparticles in a liquid flow,
where the main objectives are to maximize the num-
ber of detected particles and minimize the problems
related to overlapping. In order to evaluate and de-
termine the accuracy of the algorithms, synthetic im-
ages, of known content, have been created. They al-
low us to vary only one parameter at a time and com-
pare the results obtained at the end of the processing
with the known and imposed parameters used for the
generation of images.
2.1 Synthetic Images: Dataset
Three groups of synthetic two-dimensional (1024 ×
768 px) images have been created, where each group
consists of 40 images including 200 randomly dis-
tributed particles with different properties such as par-
ticle size, particle’s light intensity, and Gaussian noise
rate. Figure 1 shows samples of the synthetic dataset
images for which we varied one or more parameters
(e.g., size dispersion, background noise, illumination
dispersion).
2.1.1 Dataset Creation
In order to test the implementation and robustness of
our algorithm, we need to have images, especially
synthetic images that are modelled on the basis of
real-world experimental ones. These particle image
recordings are based on different characteristics like,
for instance, particle position, diameter, shape, dy-
namic intensity range, spatial density, image depth,
flow patterns, noise in the image, etc. These synthetic
images with well-defined particle locations also help
us in quantifying the quality of our algorithms by pro-
viding measurement error estimation (Ohmi and Li,
2000; Raffel et al., 2018; Mohr et al., 2019).
The synthetic images required to test our detec-
tion algorithm are generated based on the following
steps (Raffel et al., 2018; Thielicke, 2021):
• Size (height × width) and background
(black/white) of the images to be generated.
• Parameters like number N, diameter d
τ
, and flow
pattern of particles along with the type of noises
(e.g., Gaussian, Salt and Pepper, Poisson) to be
added in the images.
• Creation of particles based on their intensity, size
and centroid positions (X
0
,Y
0
,Z
0
) in the first im-
age.
Application of Particle Detection Methods to Solve Particle Overlapping Problems
85