2.3. Image processing method
All the image-based 2D measurements have been performed along thex -z plane shown in Figure 1(a). Note, unless specially
mentioned, all imaging data have been acquired during the steady-state
of the fluidization at least 5 min after the start of supplying the gas
flow. The lower frame rate data (100 fps) have been used for extracting
the expanded bed height temporal evolution matching the acquisition
frequency of pressure signals. The detailed image processing algorithms
have been reported in Gao et al .7 Briefly, the
processing steps include, (1) identifying the region of interest and
global thresholding using Otsu mehod28, (2) filtering
out sticky particles clusters on the wall due to electrostatic forces
using morphological and size filters, and (3) estimating the bed height
as the maximum height that particle pixel fraction at which height
reduces to 5%. The times series are used later for the calculation of
the mean and standard deviation of bed height, as well as the bed height
spectrum. At least 100 seconds of bed height data has been acquired for
statistically robust results. Figure 2 provides a sample convergence
test of velocity statistics for the bed height data on mean bed height
and standard deviation of bed height as a function of sample time
length, denoted as <H b>t and \(\sigma_{H_{b,t}}\). As it
shows, both the mean and the standard deviation value fluctuate
intensely for ~10 second and approaches to a constant
value as the sample time length increases.
The higher frame rate particle image series at 800 fps have been
recorded to acquire the particle velocity and wood particle orientation
statistics. The processing steps involve three phases: (a) segmentation
of particle types and background, (b) generate particle images
containing only one component, (c) perform PIV or PIV on the
corresponding image series. During phase (a), owing to the complexity of
the backlit images due to illumination and binary particle spatial
distribution, segmentation based on global or adaptive local
thresholding methods has been proved inadequate for accurately
identifying different particle types. Consequently, a machine
learning-based pixel-wise classification has been applied, following
Arganda-Carreras et al.29 The resulting probability
maps of classified pixels are then segmented by Otsu’s method. Figure 3
provides a sample image series showing the results of the steps during
wood particle extraction. As shown in Figure 3(b), the segmentation
results based on the machine learning approach successfully classify all
pixels into three categories, namely wood particles (red), LDPE
particles (green), and background (purple). Afterward, the wood particle
masks are generated (Figure 3c,d), and the geometric parameters
including, particle center, width/length, aspect ratio, and orientation
are measured (Figure 3e). Due to the inherent limitation in 2D imaging,
the 3D orientation of the particle relative to the bed central axis is
not readily measured. To eliminate the limitation, only particles
aligning almost parallel to the imaging plane (x-z plane, see Figure 1)
are sampled as statistics. This is implemented by filtering out all
particles having lengths less than 90% of the length. The detected wood
particles videos are provided as supplemental information. Based on the
particle location, wood particles are tracked undergoes particle
tracking velocimetry (PTV) to provide Lagrangian particle tracks of the
wood pellets, using an algorithm developed previously by Ouellette et
al.30. Only tracks with lengths larger than 5 frames
have been used. Next, the wood pellets are masked out in the original
image. The Eulerian velocity fields of the LDPE particles were then
computed using a particle image velocimetry (PIV) open software
PIVlab31, with a final interrogation windows size of
64 × 64 pixels with 50% overlap. Figure 4 shows the sample images
illustrating the PIV analysis. Figure 4a shows a raw image with LDPE
particles with the inserts showing the particle image of the 64 × 64
pixels interrogation window. The LDPE particles mask (Figure 4b)
generated using the abovementioned machine-learning algorithm was used
for determining the LDPE particle pixel percentage within the
interrogation window. Only interrogation windows with LDPE particle
pixel percentages larger than 10% are used for providing velocity
values of the LDPE particles as shown in Figure 4c.