Automated Detection of Primary Particles from Transmission Electron Microscope (TEM) Images of Soot Aggregates in Diesel Engine Environments.
Diesel soot particles within a diesel flame [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] or in the exhaust stream [16, 17, 18, 19, 20] show aggregate structures composed of various numbers of spherical primary particles. The soot aggregates vary not only in size but also in shape ranging from long-stretched to highly agglomerated structures. The number of primary particles in each aggregate also varies from the aggregates comprised of tens of primary particles to double-primary (dimer) aggregates or even single-primary (monomer) particles. Many studies report a correlation between the soot structures and their toxicity [21, 22, 23, 24, 25], particularly noting high reactivity of ultrafine particles to cause the potential damage of a respiratory system [26,27]. Therefore, the need is clear for a correct measurement of the size and shape of soot aggregates as well as primary particles.
The sampling of soot aggregates on a carbon-coated mesh grid and subsequent transmission electron microscope (TEM) imaging is the most popular technique to clarify soot size and structures [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]. The indexing and size measurement of soot aggregates is straightforward since the binary TEM image using a threshold-based boundary detection algorithm can provide the radius of gyration of each aggregate . However, the detection of primary particles within diesel soot aggregates poses a significant challenge due to their low contrast boundaries and the particles overlapping [8,9].
A correct measurement of the primary particle diameter ([d.sub.p]) is critical for the analysis of aggregate morphology and soot optical properties such as fractal dimension ([D.sub.f]) and dimensionless extinction coefficient ([k.sub.e]), respectively. For instance, the characteristic fractal dimension may be obtained using the following power law function :
N = [k.sub.f] ([[R.sub.g]/[[d.sub.p]]]) (1)
where N is the number of primary particles within a soot aggregate, is the fractal prefactor, [R.sub.g] is individual soot aggregate size, and [bar.d.sub.p] is the mean diameter of primary particles of each aggregate . The and [k.sub.f] can be obtained by reading the slope and the intercept of the least-square straight line of a plot ln(N) over In ([R.sub.g]/[bar.d.sub.p]). In this equation, higher [D.sub.f] of a certain soot aggregate means more compacted shapes while a lower fractal dimension indicates more stretched aggregates. The N in Eq. (1) can also be obtained using [R.sub.g] and [bar.d.sub.p] :
[mathematical expression not reproducible] (2)
where [A.sub.a] is the projection area of a soot aggregate and [A.sub.p] is the mean projection area of primary particles of the aggregate. Previous studies suggest that the empirical constant [k.sub.a] varies from 1 to 1.81 while the primary particle overlap factor [alpha] is in the range of 1.08~1.19 [29, 30, 31, 32, 33, 34]. In Eqs. (1) and (2), it is clear that the measurements for [d.sub.p] is very important to obtain the number of primary particles of each aggregate and the fractal dimension.
Another instance in which [d.sub.p] plays a key role is the estimation of the dimensionless extinction coefficient. Soot volume fraction ([f.sub.v]) is determined from the dimensional extinction coefficient data using the following relationships :
[f.sub.[upsilon]]= [K[lambda]/[k.sub.e]] and [k.sub.e] = (1 + [[rho].sub.sa])6[pi]E(m) (3)
where K is the extinction coefficient, X is the laser wavelength, [k.sub.e] is the dimensionless optical extinction coefficient, [[rho].sub.sa] is the scattering-to-absorption ratio, m is the refractive and E(m) is a function of the refractive index of soot:
E(m) = -Im ([[m.sup.2]-2/[m.sup.2]+2]) (4)
In Eq. (3), the [[rho].sub.sa] can be obtained using the Rayleigh-Debye-Gans scattering absorption theory for polydisperse fractal aggregates (RDG-PFA) [36, 37, 38]:
[mathematical expression not reproducible] (5)
[[omega].sub.p] = [2/3] [([2[pi]/[lambda]]).sup.3] ([[d.sub.p]/2]).sup.3] [F(m)/E(m)] (6)
In Eq. (5), [[omega].sub.p] is the albedo of a primary particle and F(m) is a function of the refractive index of soot:
F(m) = [|[[m.sup.2]+1/[m.sup.2]+2]|.sup.2] (7)
Equations (5) and (6) suggest that the measured [d.sub.p] together with [R.sub.g] and [D.sub.f] from Eq. (1) can provide the dimensionless extinction coefficient. Therefore, the motivation is clear for the accurate measurement of [d.sub.p].
Despite its importance, the [d.sub.p] measurement is heavily dependent upon a manual selection of primary particles from TEM images [4, 5, 6,8,9,13]. Manual detection methods require substantial time and effort to obtain large numbers of primary particles for statistically meaningful results. For biological macromolecules applications, automated particle detection from electron micrographs has been attempted based on component labelling and feature computation  or textual methods ; however, the level of false positives remains high and there is no guarantee that these methods would be directly applicable for soot particles.
Recently, Grishin et al.  developed an automated detection method for soot primaries which is based on the Circular Hough Transform (CHT) . The CHT algorithm functions by fitting circles of various sizes onto pixels that are identified as "candidate pixels" . Prior to the CHT, the Otsu threshold algorithm, rolling-ball transform (smoothing), and an edge-following algorithm (Papert's turtle approach) are used for noise reduction, boundary cleaning, and identification of the edges of the soot aggregates, respectively, as their method uses the boundary pixels as the candidate pixels. The results are a series of curve fragments, which can be fed into the CHT for the primary particle detection. Due to unexpectedly high curvature fragments, the regularisation is required but the final product shows a good agreement with manually selected primary particles.
This study presents the improvement of the CHT-based automatic detection of soot primary particles in two aspects. One is the use of Canny Edge Detection (CED)  for the processing of not only the aggregate edge pixels but also inner pixels because some primary particles are surrounded by other primary particles within each aggregate. The CED was promising because a similar method was successfully applied to the detection and sizing of subsurface spherical particles shown in noisy and low contrast TEM images . The other is the pre-processing of TEM images using image inversion and self-subtraction as well as unsharp filtering (negative Laplacian transform), which is expected to enhance the particle boundaries while suppressing noise signals. How each of image-processing parameters affects the automated primary particle detection is also discussed for soot particles sampled directly from diesel flames in both a constant-volume combustion chamber and a diesel engine. The results are compared to our previous data obtained using a manual selection method. The full Matlab code used in this study and its application to three different sets of soot TEM images from three other diesel combustion facilities are also presented.
SOOT SAMPLING EXPERIMENTS
The sampling of soot particles was conducted in five different facilities simulating diesel engine conditions, of which three are optically-accessible, pre-burn-type constant-volume combustion chambers (CVC) available at Meiji University [3,4,6, 7, 8,11,12,14,15], Sandia National Laboratories [2,4,35,46], and IFPEN [47, 48, 49], and the other two are diesel engines housed in the University of New South Wales (UNSW) [9,10,13] and Lund University . In this study, soot samples from the Meiji CVC and the UNSW diesel engine were used for the development of the new particle detection code and its performance evaluation for various image-processing parameters. The two facilities were selected because extensive datasets of the soot primary particles are available from the manual selection methods, which can be compared to the new data from the automatic method. The ambient gas and injection conditions of these two facilities were also very close. Soot samples from the other three facilities were used to demonstrate the efficacy of the new code and selected image-processing parameters, and to provide example cases for various soot samples. Details about the used facilities and operating procedures for soot sampling experiments are found in our previous studies [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 46, 47, 48, 49, 50, 51] and thus only a brief summary is provided in this section.
MEIJI UNIVERSITY CONSTANT-VOLUME COMBUSTION CHAMBER
Figure 1 shows the schematic of the Meiji CVC (bottom-left), soot particles sampling (top-left), and a TEM image of sampled soot particles (top-middle). The ambient gas and fuel injection conditions are listed in Table 1. The chamber was operated so that fixed ambient gas conditions of 2.5 MPa, 940 K, and 21% [O.sub.2] concentration are created at the time of fuel injection and throughout the diesel combustion event using the spark ignition of ultra-lean mixture of acetylene, oxygen and nitrogen. A conventional diesel fuel (JIS#2) was injected via an axially-drilled, 140-[mu]m hole of a common-rail injector at fixed rail pressure of 80 MPa for the injection duration of 2.5 ms, which corresponds to 10.3 mg of fuel mass per injection.
The soot particles were sampled at 70 mm from the nozzle, which was about 65 mm downstream of the flame base at 5 mm from the nozzle and where the highest amount of soot is measured [3,4,7,11,15]. The thermophoretic force (i.e. positive thermal diffusion) between high-temperature soot particles within the flame and a cold (373 K) carbon-coated copper mesh was used for the sampling, which is believed to quench the soot reactions instantly . The sampling probe placed in parallel to the sooting diesel flame (see Fig. 1) has a small 3-mm-wide and 1.5-mm-deep opening at the tip, which provides a passage for the soot-laden gas flow while protecting the 3-mm copper grid from excess heat, consistent with Refs. [2,4].
A transmission electron microscope (JEOL JEM-2100F) readily available at Meiji University was used to take high-magnification soot images. The specified point resolution of the microscope is 0.19 nm and the electron acceleration voltage is 200 keV. All images were digitally recorded using a CCD camera (Gatan UltraScan 1000) with an image resolution of 2048 by 2048. An example image is shown at the top-middle of Fig. 1, which shows various soot aggregate structures. A selected aggregate (illustrated by a red square) will be used for the discussion of the automated detection of soot primary particles in the following sections.
Sandia National Laboratories and IFPEN Constant-Volume Combustion Chambers
The other two CVCs at Sandia [2,4,35,46] and IFPEN [47, 48, 49] operate in a very similar manner in terms of the pre-burn of ultra-lean mixture to create high-temperature and high-pressure ambient gas conditions of diesel engines as well as the use of an axially-drilled single-hole nozzle in a common-rail injector. As summarised in Table 2 and 3, however, the Sandia CVC was run at higher ambient gas pressure condition (6.7 MPa), lower ambient [O.sub.2] concentration (15%), and higher injection pressure (150 MPa) than those of the Meiji CVC. In comparison, the IFPEN CVC had the same 21% [O.sub.2] concentration as of the Meiji CVC but higher ambient gas pressure of 6 MPa and higher injection pressure of 150 MPa, similar to the Sandia CVC (i.e. [O.sub.2] concentration variants of the ECN Spray A target conditions ). The fuel used for the IFPEN (n-dodecane) was different to the Meiji and the Sandia experiments (conventional diesel fuel). Variations in ambient gas and fuel injection condition are well known to affect the soot particles [4,13,15,20]. Therefore, the soot particles sampled from these CVCs provide a good opportunity to test the new automated detection method for a wide variety of soot sizes and structures.
The soot sampling experiments were conducted using a similar method as in the Meiji CVC (details are found in our previous work [2,49]). Regarding TEM imaging, the soot samples (TEM grids) from the IFPEN CVC were sent to Meiji University for processing and thus the same method and set values of the Meiji TEM were used. However, the Sandia samples were processed at an in-house facility [2,4]. Obviously, different microscopes used by different operators can lead to variations in the resulting TEM images. To address this issue, the new code implements a background correction step prior to the edge detection and primary particle recognition as will be discussed later in the Algorithms section.
UNSW Optical Diesel Engine
Figure 1 (right) shows the schematic of the UNSW single-cylinder optical diesel engine used for in-flame soot sampling experiments via thermophoresis. Specifications of the engine and its operating conditions are listed in Table 4. The engine has a displacement volume of 498 cc with a small 83-mm bore and 92-mm stroke, typical of a light-duty diesel. To create thermally stable engine conditions, the wall temperature was held constant at 363 K by circulating heated water to the engine head, cylinder liner and engine block. All experiments were conducted at fixed engine speed of 1200 revolutions per minute (rpm) controlled by an electric motor.
The used fuel was a conventional ultra-low-sulphur diesel with cetane number of 51. The fuel injection system was composed of a Bosch second-generation common-rail and a 134-[micro]m single-hole injector. At fixed rail pressure of 70 MPa, the injection occurred for 2.34 ms, which corresponds to 9 mg of fuel mass per injection. It should be noted that these injection conditions are very similar to the Meiji CVC conditions (see Table 1). The injection timing was fixed at--7 crank angle degrees after top dead centre ([degrees]CA aTDC) and a 10-skip-firing strategy (injection every 10th cycle) was used to minimise cyclic dispersion caused by residual gas from previous firing cycles.
For the soot sampling, a similar method to the Meiji soot sampling was used involving a carbon-coated copper grid with 400 meshes and a restrictive passage (1-mm hole) at the sampling probe tip. As shown in Fig. 1(right), however, the sampling probe was not parallel to the sooting diesel flame but was placed within a 30-mm-wide bowl-rim cut-out region. This bowl-rim cut-out was required to avoid a potential crash between the stationary sampling probe and the fast-moving piston . The TEM grid was located about 62 mm from the nozzle along the jet travelling trajectory. A transmission electron microscope (JEOL 1400) with a point resolution of 0.38 nm and an accelerating voltage of 120 kV was used to obtain high-magnification soot images, which was available at UNSW. A CCD camera with a resolution of 11 mega pixels was used to digitise soot particle images. An example image is shown at the bottom-middle of Fig. 1. Similar to the Meiji CVC image, a selected aggregate for the discussion of the automated detection method is illustrated by a red rectangle.
Lund University Diesel Engine
The Lund University diesel engine  is an in-line six-cylinder 12.78-litre engine with a large bore and stroke of 131 and 158 mm, respectively for heavy-duty applications. The engine specifications and operating conditions are listed in Table 5. The multi-cylinder engine was used for the soot particles sampling in the exhaust, which is complementary with the in-flame sampling conducted in the UNSW's light-duty diesel engine. The engine was run at 1200 rpm with 43.3% of a recirculation ratio of the exhaust gas resulting in 10.4% oxygen concentration in the intake air. Similar to the UNSW engine, a conventional ultra-low-sulphur diesel with cetane number of 51 was injected using a six-hole unit injector at maximum injection pressure of 200 MPa. The fuel injection was adjusted for the estimated equivalence ratio of 0.645, corresponding to 1 MPa indicated mean effective pressure (IMEP) conditions at fixed CA50 (crank angle location for 50% heat release) of 3[degrees]CA aTDC.
The Lund diesel engine was used for exhaust soot sampling, similar to our previous work . The exhaust gas was diluted using a two-stage ejector dilution system (Dekati) performing 1:8 dilution with 431-K filtered air and additional 1:15 dilution with air at room temperature. Between the two dilution processes, the gas was heated to 573 K for about 2 seconds to remove low volatility organics from the particles. This setup corresponds to the European PMP sampling standard, which is developed to avoid nucleation of volatile nanoparticles and condensation of low volatility organics on the soot aggregates. The soot aggregates in the diluted exhaust gas were sampled using an electrostatic precipitator (Nanometer Aerosol Sampler Model 3089, TSI) at a 9.6-kV voltage and 1-lpm flow rate conditions. The sampled soot aggregates were imaged in the Meiji transmission electron microscope.
AUTOMATED DETECTION METHOD
The new method for automated particle detection involves four steps including image inversion and self-subtraction, unsharp masking based on the negative Laplacian transform [52,54], Canny Edge Detection (CED) , and Circular Hough Transform (CHT) [42,43]. The full code written on Matlab (version 2012a or higher) with an image processing toolbox is provided in the Appendix section. The overall procedure for the image processing is shown in Fig. 2. A digitised TEM image contains about 20~30 soot aggregates. A selected soot aggregate (see Fig. 1) from the UNSW diesel engine data is used as an example in this figure.
The background of TEM images varies sample by sample and image by image due to the differences in the carbon film and the magnification/intensity setting of TEM imaging. This was corrected by applying the combination of image inversion and self-subtraction. In TEM images, the soot particle regions show lower image counts than the carbon film area (e.g. the top-left of Fig. 2). The image inversion through the subtraction of the original image count from the maximum image count of the TEM image results in higher image counts in the soot particle region for easier threshold determination in the following steps. Then, the original TEM image is subtracted from the inverted image. This self-subtraction procedure effectively increases the image contrast because brighter soot particle regions in the inverted TEM image undergo less image count reduction compared to darker carbon film regions. Some pixels with negative image counts as a result of this self-subtraction are set to zero value, leading to a black background, as shown in Fig. 2 (top-middle).
The next step is a 3-by-3 unsharp filtering to enhance the edges of soot primary particles. The unsharp filter is from the negative of the Laplacian filter with a selected control parameter for the Laplacian shape [52,54]. The resulting image is shown in Fig. 2(top-right).
The sharpened image works as an input to the Canny Edge Detection (CED) utilising a multi-stage algorithm to detect a wide range of edges [44,45]. The CED consists of five steps including the Gaussian filter (smoothing) for noise reduction, the finding of image intensity gradients, the reduction of spurious response to edge detection by applying non-maximum suppression, the detection of double thresholds, and finally the hysteresis based edge tracking. The final step suppresses all the other edges that are weak and not connected to strong edges. The resulting image is shown at the bottom-right of Fig. 2 with many edges detected at the aggregate boundaries and within the aggregate.
Finally, the soot primary particles are selected by the Circular Hough Transform (CHT) of the Canny edge image. The CHT implements a voting process to map image edge points into manifolds in an appropriately defined parameter space . Peaks in the space correspond to the parameters of detected curves. The CHT requires two input parameters of the circle size range (minimum and maximum radius) and sensitivity. The primary particles detected by the CHT are overlaid on the Canny edge image as shown at the bottom-middle of Fig. 2. The same primary particle boundaries are also overlaid on the original TEM image for the selected soot aggregate (bottom-left of Fig. 2).
Comparison Between Manual and Automatic Methods
One way to validate the newly developed particle detection code is its comparison to manually selected soot primary particles. The manual method is obviously arbitrary as it depends solely on one's visual inspection which can vary person by person depending on their level of understanding of soot structures. However, our previous work suggested that the selection process conducted by experienced operators (i.e. those with a clear understanding of soot primaries and their appearance on TEM images) can limit the fluctuations within [+ or -]3.5% . Also, it was found that the size distribution of primary particles using a large sample volume with over 2500 selected particles shows near Gaussian distributions [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15], adding credibility to the mean value.
In Fig. 3(left), examples of the manual selection methods are shown in green rectangle boxes for selected aggregates from the Meiji CVC (noted "CVC" at the top) and UNSW diesel engine (noted "Engine" at the bottom). The images show that many circular primary particles stack up together to form an agglomerate of complex shape. Measurements for the diameter of these primaries were manually selected and overlaid on the aggregate image as green lines connecting one end to the other of primary particles. On the right of Fig. 3, primary particles from the automated detection method are overlaid on the same aggregates.
Some mismatches are seen in Fig. 3. For example, the automated method cannot detect a few primary particles such as primary 1 and 3. However, there are opposite cases that the automated method finds primary particles that are ignored in the manual method (e.g. primary 2 and 4). Overall, a good agreement is seen between the automated and manual methods while the automated method detects less primary particles than that of the manual selection. This is simply because the CHT requires a clear circular object whereas the manual selection is possible for incomplete structures depending on one's shape recognition.
Figure 4(top) shows the size distribution of 140 primary particles from the manual method together with the automatically detected 89 primaries. The results are from the selected TEM image of Fig. 3 (the Meiji sample [4,8]). At the bottom of Fig. 4, the results from the UNSW diesel engine are plotted for 6606 and 1636 primary particles from the manual and automated method, respectively. A total of five TEM images were processed [9,13] including the selected TEM image of Fig. 3 (the UNSW sample) to evaluate the performance of the new code for larger sample sizes. The size distributions of primary particles display overall a good match and the mean [d.sub.p] show identical values of 17.9 and 16.4 nm for the CVC and the engine samples, respectively. It is also noticeable that the mean [d.sub.p] is only 1.5 nm different between the Meiji CVC and the UNSW diesel engine cases, likely due to similar ambient gas and fuel injection conditions (see Tables 1 and 2). In comparison, this is much less than a difference of 5~6 nm caused by a closer sampling location of 40 mm from the nozzle  or higher injection pressure of 160 MPa .
The results from the automated detection method in Figs. 3 and 4 were obtained using many input parameters for the image processing. These parameters are listed in Table 6 together with other set values required for the implementation of Matlab code provided in the Appendix section. First of all, the TEM scale (nm per pixel) varies depending on the magnification setting in the used TEM and a resolution of the CCD camera. The maximum image count is either the image count of the carbon film area (as in the CVC case for a 32-bit image) or the maximum value of the digitised image (as in the Engine case for an 8-bit image). In the Matlab code, median filtering is included to consider TEM images with "salt and pepper" noises. However, none of the TEM images in this study include high noises and thus the median filtering set value was fixed at 1 (for 1 pixel by 1 pixel), meaning no filtering.
There are five image-processing parameters affecting the automated detection of primary particles as shown in Table 6. There include 1) the self-subtraction level, 2) the negative Laplacian shape parameter, 3) the maximum [d.sub.p], 4) the minimum [d.sub.p], and 5) the CHT sensitivity. Each of these parameters is evaluated in the following sections using the same example aggregates of Fig. 3. The size distributions of primary particles are also plotted for various set values from both the manual and automatic methods. When the parameter of interest is varied, the other four parameters are fixed at a value specified in Table 6.
The first step of the image pre-processing is the image inversion and self-subtraction (see Fig. 2), which effectively enhances the contrast of the soot region. In this step, the self-subtraction parameter is multiplied to the original image (see step 1 of the Matlab code in the Appendix section), which can be varied depending on the contrast of the original TEM image. How the subtraction level affects the primary particles detection is shown in Fig. 5. The same soot aggregates of Figs. 3 are used once again as an example aggregate. For each aggregate, three self-subtraction levels are presented for the optimised value shown in the middle (red font) and 0.1 increments lower and higher than the optimised value.
Figure 5(left) shows that the TEM image from the CVC is overall darker and has a less difference in image count between the soot and carbon file regions compared to those for the TEM images from the engine. This results in a higher subtraction level for the CVC image so that the signals in the soot region stand out over the carbon film region. For the soot primary particle detection, it is seen that a lower self-subtraction level detects false large primaries whereas a higher self-subtraction level tends to select false small primaries.
The size distributions of all primary particles (i.e. thousands of primaries as in Fig. 4) from both the manual and automatic methods are shown for various self-subtraction levels in Fig. 5(right). In the figure, probability density functions (pdf) of primary particle size ([d.sub.p]) are shown together with the mean [d.sub.p] and error range corresponding to 2[sigma]/ [square root of n] where [sigma] is a standard deviation of n primaries (95% confidence). The pdf is preferred over the number count (e.g. Fig. 4) due to lower sample size of the automatic method. The results tend to show increased pdfs at a lower [d.sub.p] range with an increasing self-subtraction level. By contrast, higher pdfs are found at a high [d.sub.p] range when a lower subtraction level is used. If the optimised self-subtraction level of 1.2 and 0.8 are used for the CVC and Engine sample, respectively, the pdfs exhibit a good agreement between the manual and automatic methods. Also, the mean [d.sub.p] show a perfect match for the optimised self-subtraction level.
Negative Laplacian Shape Parameter
The inverted and self-subtracted TEM image requires image sharpening to improve edge detection performance in the following step. The standard tool of choice for sharpening is unsharp filtering . The unsharp filter based on the negative of the Laplacian filter is controlled by a parameter determining the shape of the Laplacian . In this study, the unsharp filtering was implemented using a Matlab function called "fspecial" with "unsharp" type and a modifying parameter "alpha" (see the Appendix section). The resulting images are shown in Fig. 6 for various alpha values.
A noticeable trend in the images of Fig. 6(left) is a mild shrinkage of detected primary particles when the alpha increases from 0 to 0.1 for the CVC images because the 0.1 alpha effectively reduces the over-estimated primary particles. For the engine images, the 0.1 alpha also performs better than the no unsharp filtering case (i.e. alpha = 0) with the increased number of circles better matching the primaries within the aggregate. When the alpha increases further to 0.5, both the CVC and the engine images start to lose circles that are clearly primary particles. Figure 6(right) shows the [d.sub.p] pdfs for both the CVC (top) and the engine (bottom) samples. It appears that the pdfs do not vary significantly with the alpha variations. The mean [d.sub.p] also shows a good agreement with the results from the manual method for the alpha of 0.1, consistent with the example images on the left.
The image processing step next to the unsharp filtering is the Canny Edge Detection (CED). It was noted that the CED implemented in our code has no control parameter and thus the pre-processing through the image inversion, self-subtraction, and unsharp filtering essentially controls the results for Canny edges.
Radii Range for the Circular Hough Transform
The Circular Hough Transform (CHT) requires two input parameters of a size range (maximum and minimum circle size) and a sensitivity value. In our image-processing code, the CHT was implemented using "imfindcircles", a Matlab function for the circular object detection, with minimum and the maximum pixel values for circles and the sensitivity (between 0 and 1) as input parameters. For the two-stage based CHT used in this study, primary particles are explicitly estimated utilising the estimated circle centres along with image information, which is based on computing radial histograms [43,45].
Figure 7 shows the soot aggregate images for various maximum diameters ([d.sup.max.sub.p]) of the primary particles. A first noticeable trend from this figure is that for smaller [d.sub.p.supmax] (e.g. 20 nm for the engine case), many primary particles evident with circular structures cannot be detected due to the size restriction. When higher [d.sup.max.sub.p] is applied (e.g. 43.5 and 80 nm for the CVC and the engine case, respectively), however, it is seen that some of small primary particles are ignored because the radii range becomes too wide. The observed trends suggest there should be an optimised [d.sup.max.sub.p] in between these values, e.g. 41.7 and 60 nm for the CVC and the engine case, respectively. However, the pdfs in Fig. 7(right) show minor variations and the mean [d.sub.p] appears to be very similar with the results from the manual method if the very small [d.sup.max.sub.p] of 20 nm is excluded. This was because the chance of primary diameters being higher than 40 nm is very low for both the CVC and the Engine samples.
While a large difference of [d.sup.max.sub.p] (e.g. 20 nm increments) makes minor variations in the pdfs, a small change in the minimum [d.sub.p] ([d.sup.min.sub.p]) appears to make a significant impact on both the primary particle detection and the pdf. Figure 8(left) shows the results for various [d.sup.min.sub.p]. It is clearly seen that the [d.sup.max.sub.p] of about 8 nm works well for both the CVC and the engine samples. It is also evident that with only about 4-nm change in [d.sup.max.sub.p], a lot of false primary particles are detected. For lower [d.sup.min.sub.p], very small and large false primaries are selected because the radii range is too wide. By contrast, many primaries are missed out for higher [d.sup.min.sub.p] due to the size restriction.
The size distributions shown in Fig. 8(right) also exhibit a dramatic change with much higher pdfs in the small [d.sub.p] range for lower [d.sup.min.sub.p] and vice versa. The mean [d.sub.p] shows a good agreement when an optimised [d.sup.min.sub.p] of about 8 nm is used. An important note to make at this point is that among four image-processing parameters evaluated so far (Figs. 5, 6, 7, 8), the [d.sup.min.sub.p] shows the most significant impact on the primary particles detection.
The Circular Hough Transform (CHT) Sensitivity
In the previous studies [41,45], the significance of the CHT sensitivity on circular object detection was well demonstrated. The higher sensitivity would be preferred for the selection of very small circles; however, the detection of false circles also increases. The use of a combination of noise reduction and boundary cleaning algorithms , median filtering and Laplacian filtering , or the self-subtraction and unsharp filtering of this study have resolved this issue to some extent but not completely. Therefore a careful selection of the Hough sensitivity is required.
Figure 9 shows a well expected trend of increasing false detection of small circles with an increasing sensitivity value. By contrast, the low sensitivity tends to ignore many primary particles. The size distribution in Fig. 9(right) displays increasing pdfs in the low [d.sub.p] range with the increasing Hough sensitivity. The mean [d.sub.p] also shows measurable variations with the sensitivity change while the best agreement with the manual processing results is found at 0.79 and 0.75 for the CVC and the Engine samples, respectively.
In summary, the results in this section suggest that the automated primary particle detection is not much sensitive to the inversion and self-subtraction, unsharp filtering, and the maximum [d.sub.p] value; however, extra care should be taken in selecting the minimum [d.sub.p] and the Hough sensitivity.
The optimised values for the image-processing parameters in Figs. 5, 6, 7, 8, 9 and Table 6 are similar for the Meiji CVC and the UNSW diesel engine samples, if differences in the original TEM images were considered carefully by varying the maximum image count and self-subtraction level. However, this consistency might be simply due to similar injection and ambient gas conditions between the two experiments. Therefore, this section shows evaluation of the new code for three more sets of soot TEM images obtained not only at three different injection and ambient gas conditions but also three different combustion facilities including two pre-burn type CVCs and a heavy-duty diesel engine (exhaust).
Sandia CVC Data
Figure 10 shows the application of the automated particle detection code for a soot TEM image obtained from a constant-volume combustion chamber of Sandia National Laboratories. A first noticeable trend from the top-left image for the particle boundaries overlaid on the TEM image is a lot more soot aggregates compared to the TEM images of the Meiji CVC or the UNSW diesel engine samples (see Fig. 3). This was largely due to a lower magnification setting of the microscope (see the relative size of the 200-nm scale bar) but to some extent, higher soot formation occurred at lower oxygen concentration of 15% at 1000-K ambient gas temperature. It is seen that the primary particles detection was successful as evidenced by lots of red circles illustrating the processed primaries. Also, the example aggregates at the top-right of Fig. 10 shows good results with most of the aggregate areas being covered by the primary particles.
The set values of the image-processing parameters used to obtain the results in Fig. 10 are summarised in Table 7(second column). As shown at the top-right of Fig. 10, the in-house microscope used at Sandia National Laboratories provides an 8-bit image, similar to the UNSW TEM (see Fig. 3) but the image contrast appears to be lower with less image count differences between the soot and carbon film regions. Compared to the UNSW set values (see Table 6), this required a lower self-subtraction level of 0.5 as the higher level would select false small primaries (e.g. Fig. 5). Other than the self-subtraction level (red font to denote the change), table 7 shows the same values for the image-processing parameters as of the UNSW sample to obtain the good agreement in Fig. 10. It was noted that the high sensitivity parameters (i.e. the minimum [d.sub.p] and Hough sensitivity) did not require an adjustment. Instead, the pre-processing parameter to increase the TEM image contrast was enough to obtain good results despite markedly different ambient gas and injection conditions. The size distributions of primary particles obtained from all aggregates of the top-right TEM image are shown at the bottom of Fig. 10. A total of 3019 primaries were processed from a single TEM image, which are mostly 10~20 nm in diameter. The mean [d.sub.p] is measured at 15 nm, lower than the Meiji or the UNSW samples (Fig. 4) likely due to higher injection pressure of 150 MPa .
IFPEN CVC Data
As mentioned previously, the IFPEN soot samples were sent to Meiji University for TEM imaging and thus the original TEM image shown at the top-left of Fig. 11 is similar to the Meiji sample image (Fig. 3) with overall lower brightness (low image counts) and lower image contrast. The Meiji University TEM provides a 32-bit image and thus requires an adjustment of the maximum image count depending on the image count in the carbon film region. Therefore, the same values for the image- processing parameters but the maximum image count were used for the processing of the IFPEN sample, as shown in Table 7(third column). It is noted that a higher maximum image count of 18000 than that of the Meiji sample was selected considering very dark carbon film region.
The resulting image is shown in Fig. 11(top-left) for the primary particle boundaries overlaid on the original TEM image. Compared to the Sandia image (Fig. 10), the number of aggregates is much lower even if the higher TEM magnification is considered. It was noted that the injection pressure was high 150 MPa as the Sandia experiments but higher oxygen concentration of 21% as well as the use of a low-sooting n-dodecane fuel led to reduced soot formation. Also shown at the top-right are two selected aggregates as an example for primary particles detected within the aggregates. Once again, the application of the new code is successful with a good agreement between the detected primaries and the visual inspection. Figure 11(bottom) shows that a total of 64 primary particles were processed with the mean [d.sub.p] of 11 nm. This mean [d.sub.p] is the lowest among the samples of this study likely due to the use of n-dodecane, a low-sooting fuel with no aromatics.
Lund Diesel Engine Data
The Lund University soot samples cover the practical end of the research spectrum because a multi-hole injector was used in a heavy-duty diesel engine for realistic jet-jet interactions effect on soot. The exhaust soot aggregates sampled from the Lund University diesel engine are products of complex soot formation and oxidation occurred inside the engine cylinder and thus the in-flame sampling as in the other four combustion facilities of the present study would be preferred for the clarification of soot processes. However, the exhaust soot aggregates provide valuable information about the soot morphology and are directly relevant to their impact on the environment and human health.
Similar to the IFPEN samples, the soot samples of the Lund University diesel engine were sent to Meiji University for TEM imaging. The resulting image is shown in Fig. 12. A first noticeable trend from the soot TEM images at the top-left of Fig. 12 is that the number of soot aggregates is very low. It was because a low-temperature combustion (LTC) strategy was implemented by applying a high ratio of exhaust gas recirculation (EGR) to achieve very low soot levels in the exhaust [50,52]. Figure 12 shows that the TEM images are overall brighter but the image contrast (i.e. the difference between the soot and carbon file regions) is even lower than that of the IFPEN sample. This required a higher maximum image count of 19000 as shown in Table 7. Other than this change, the same values of the image-processing parameters as for the IFPEN as well as the Meiji samples were enough to obtain a good agreement between the automatically processed primary particles and the visual inspection.
Figure 12(bottom) shows that a total of 69 primary particles were selected and the size distributions of detected primary particles suggest high tendency to have small primary particles (~ 10 nm) within the aggregates. The mean [d.sub.p] is measured at 13 nm.
A new image-processing code for automated detection of soot primary particles has been developed based on the Canny Edge Detection (CED) and Circular Hough Transform (CHT). Soot particles were sampled from diesel flames in four different combustion facilities including three constant-volume combustion chambers (CVC) and an optical diesel engine. The soot aggregates sampling was conducted via thermophoresis (i.e. positive thermal diffusion) between hot soot and a cold carbon film of the mesh grid, which was imaged using a transmission electron microscope (TEM). The exhaust soot sampling was also conducted in a multi-cylinder heavy-duty diesel engine using a dilution system and electrostatic precipitator. The TEM images were pre-processed based on the image inversion and self-subtraction and the unsharp filtering for reduced noise and enhanced edges. Results from the automatic detection method are compared to the manual method for various CVC or engine operating conditions. Not only the visual inspection of aggregate images but also the size distributions of primary particles were compared to find optimised image-processing parameters. Major findings of this study are summarised as follows:
* Five image-processing parameters affecting the automated particle detection are identified including the self-subtraction level, negative Laplacian shape parameter, maximum primary particle diameter, minimum primary diameter, and Hough sensitivity. The automatic method shows much higher dependency on the minimum primary particle diameter and Hough sensitivity than the other three parameters.
* Optimised set values for all five image-processing parameters can be found, which shows a good agreement with the results from the manual method.
* The self-subtraction level is determined by the brightness and contrast of original TEM images. In the tested conditions of this study, the self-subtraction level varies between 0.5 and 1.2.
* The negative Laplacian shape parameter of 0.1 is good enough to improve the performance of automated particle detection. Higher sharpening than alpha = 0.1 leads to decreased sensitivity for the particle detection.
* The significance of the minimum diameter value for the CHT is very high. By contrast, the maximum diameter is much less sensitive to the automated particle detection than that of the minimum diameter. In the tested condition of this study, the maximum diameter of 40 nm or higher performs well while the minimum diameter of 8~8.3 nm results in a good agreement with the manual method.
* The Hough sensitivity is optimised at 0.75 ~0.79 in the tested conditions of this study.
The support for this study was provided by the Japan Society for the Promotion of Science (JSPS) and the Australian Academy of Science via the Research Visit to Japan fellowship. The soot sampling experiments at UNSW were funded by the Australian Research Council via Discovery Project.
[1.] Li, Z., Song, C., Song, J., Lv, G. et al., "Evolution of the Nanostructure, Fractal Dimension and Size of In-Cylinder Soot during Diesel Combustion Process." Combust. Flame 158(8):1624-1630, 2011.
[2.] Kook, S. and Pickett, L.M., "Soot Volume Fraction and Morphology of Conventional and Surrogate Jet Fuel Sprays at 1000-K and 6.7-MPa Ambient Conditions," Proc. Combust. Inst. 33:2911-2918, 2011.
[3.] Kondo, K., Yamaguchi, T., Nishigai, H., Takano, S. et al., "High-Resolution Transmission Electron Microscopy of Soot Directly Sampled at Different Axial Locations in Diesel Spray Flame," SAE Technical Paper 2011-24-0068, 2011, doi:10.4271/2011-24-0068.
[4.] Yamaguchi, T., Kondo, K., Nishigai, H., Takano, S. et al., "Direct Sampling, TEM Analysis and Optical Measurement of Soot Particles at Different Axial Locations in a Transient Spray Flame," SAE Int. J. Fuels Lubr. 5(1):316-328, 2012, doi:10.4271/2011-01-2051.
[5.] Kook, S. and Pickett, L., "Soot Volume Fraction and Morphology of Conventional, Fischer-Tropsch, Coal-Derived, and Surrogate Fuel at Diesel Conditions," SAE Int. J. Fuels Lubr. 5(2):647-664, 2012, doi:10.4271/2012-01-0678.
[6.] Aizawa, T., Nishigai, H., Kondo, K., Yamaguchi, T. et al., "Transmission Electron Microscopy of Soot Particles Directly Sampled in Diesel Spray Flame - A Comparison between US#2 and Biodiesel Soot," SAE Int. J. Fuels Lubr. 5(2):665-673, 2012, doi:10.4271/2012-01-0695.
[7.] Sakai, M., Iguma, H., Kondo, K., and Aizawa, T., "Nanostructure Analysis of Primary Soot Particles Directly Sampled in Diesel Spray Flame via HRTEM," SAE Technical Paper 2012-01-1722, 2012, " doi:10.4271/2012-01-1722.
[8.] Kondo, K., Aizawa, T., Kook, S., and Pickett, L., "Uncertainty in Sampling and TEM Analysis of Soot Particles in Diesel Spray Flame," SAE Technical Paper 2013-01-0908, 2013, doi:10.4271/2013-01-0908.
[9.] Kook, S., Zhang, R., Szeto, K., Pickett, L. et al., "In-Flame Soot Sampling and Particle Analysis in a Diesel Engine," SAE Int. J. Fuels Lubr. 6(1):80-97, 2013, doi:10.4271/2013-01-0912.
[10.] Zhang, R., Szeto, K., and Kook, S., "Size Distribution and Structure of Wall-Deposited Soot Particles in an Automotive-Size Diesel Engine," SAE Int. J. Fuels Lubr. 6(3):605-614, 2013, doi:10.4271/2013-01-2534.
[11.] Kuribayashi, M., Ishizuka, Y., and Aizawa, T., "Sizing of Soot Particles in Diesel Spray Flame - A Qualitative Comparison between TEM Analysis and LII/Scattering Laser Measurements," SAE Int. J. Fuels Lubr. 6(3):641-650, 2013, doi:10.4271/2013-01-2576.
[12.] Okabe, K., Sakai, M., Mizutani, Y., and Aizawa, T., "Aromatic Additive Effect on Soot Formation and Oxidation in Fischer-Tropsch Diesel (FTD) Spray Flame - Morphology and Nanostructure Analysis of InFlame Soot Particles via HRTEM-," SAE Int. J. Fuels Lubr. 6(3):807-816, 2013, doi:10.4271/2013-01-2681.
[13.] Zhang, R. and Kook, S., "Influence of Fuel Injection Timing and Pressure on In-Flame Soot Particles in an Automotive-Size Diesel Engine." Environ. Sci. Technol." 48(14):8243-8250, 2014.
[14.] Kondo, K., Takahashi, J., and Aizawa, T., "Morphology Analysis of Wall-Deposited Diesel Soot Particles via Transmission Electron Microscope," SAE Int. J. Fuels Lubr. 7(3):683-692, 2014, doi:10.4271/2014-01-2637.
[15.] Kuribayashi, M., Mizutani, Y., Ishizuka, Y., Taki, N. et al., "Effects of Ambient Oxygen Concentration on Soot Processes in Diesel Spray Flame - A Qualitative Comparison between TEM Analysis and LII/Scattering Laser Measurements," SAE Int. J. Fuels Lubr. 7(3):693-703, 2014, doi:10.4271/2014-01-2642.
[16.] Ishiguro, T., Suzuki, N., Fujitani, Y., and Morimoto, H., "Microstructural Changes of Diesel Soot during Oxidation," Combust. Flame 85:1-6, 1991.
[17.] Ishiguro, T., Takatori, Y., and Akihama, K., "Microstructure of Diesel Soot Particles Probed by Electron Microscopy: First Observation of Inner Core and Outer Shell," Combust. Flame 108:231-234, 1997.
[18.] Song J., Cheenkachorn, K., Wang J., Perez J., Boehman, A.L., Young, P.J., and Waller, F.J., "Effect of Oxygenated Fuel on Combustion and Emissions in a Light-Duty Turbo Diesel Engine," Energy Fuels 16:294301, 2002.
[19.] Lee, K.O., Cole, R., Sekar, R., Choi, M.Y., Kang, J., Bae, C., and Shin, H.D., "Morphological Investigation of the Microstructure, Dimensions, and Fractal Geometry of Diesel Particulates," Proc. Combust. Inst. 29:647-653, 2002.
[20.] Zhu, J., Lee, K.O., Yozgatligil, A., and Choi, M.Y., "Effects of Engine Operating Conditions on Morphology, Microstructure, and Fractal Geometry of Light-Duty Diesel Engine Particles," Proc. Combust. Inst. 30:2781-2789, 2005.
[21.] Muzyka, V., Veimer, S., Schmidt, N., "On the Carcinogenic Risk Evaluation of Diesel Exhaust: Benzene in Airborne Particles and Alterations of Heme Metabolism in Lymphocytes as Markers of Exposure," Sci. Total Environ. 217(1-2):103-111, 1998
[22.] Lighty, J.S., Veranth, J.M., Sarofim, A.F., "Combustion Aerosols: Factors Governing Their Size and Composition and Implications to Human Health," J. Air Waste Manage. Assoc. 50:1565-1618, 2000.
[23.] Zhang, R., Khalizov, A.F., Pagels, J., Zhang, D. et al. "Variability in Morphology, Hygroscopicity, and Optical Properties of Soot Aerosols during Atmospheric Processing," Proc. Natl. Acad. Sci. U.S.A. 105(30):10291-10296, 2008.
[24.] Broday, D.M. and Rosenzweig, R., "Deposition of Fractal-Like Soot Aggregates in the Human Respiratory Tract," J. Aerosol Sci. 42(6):372-386, 2011.
[25.] Benbrahim-Tallaa, L., Baan, R.A., Grosse, Y., Lauby-Secretan, B. et al., "Carcinogenicity of Diesel-Engine and Gasoline-Engine Exhausts and Some Nitroarenes," Lancet Onco. 13 (Article):663-664, 2013.
[26.] Su, D. S., Serafino, A., Muller, J-O., Jentoft, R.E., et al. "Cytotoxicity and Inflammatory Potential of Soot Particles of Low-Emission Diesel Engines." Environ. Sci. Technol. 42(5):1761-1765, 2008.
[27.] Frank, B., Schlogl, R., Su, D.S., "Diesel Soot Toxification." Environ. Sci. Technol. 47(7):3026-3027, 2013.
[28.] Sorensen, C.M. and Roberts, G.C., "The Prefactor of Fractal Aggregates," J.Colloid Interface Sci. 186(2):447-452, 1997.
[29.] Samson, R.J., Mulholland, G.W., and Gentry J.W., "Structural Analysis of Soot Agglomerates," Langmuir 3:272-281, 1987.
[30.] Megaridis, C.M. and Dobbins, R.A., "Morphological Description of Flame-Generated Materials," Combust. Sci. Technol. 71(1-3):95-109, 1990.
[31.] Koylu; U.O., Xing, Y., and Rosner, D.E., "Fractal Morphology Analysis of Combustion-Generated Aggregates Using Angular Light Scattering and Electron Microscope Images," Langmuir 11(12):4848-4854, 1995.
[32.] Koylu; U.O., Faeth, G.M., Farias, T.L., and Carvalho, M.G., "Fractal and Projected Structure Properties of Soot Aggregates," Combust. Flame 100(4):621-633, 1995.
[33.] Oh, C. and Sorensen, C.M., "The Effect of Overlap between Monomers on the Determination of Fractal Cluster Morphology," J.Colloid Interface Sci. 193(1):17-25, 1997.
[34.] Brasil, A.M., Farias, T.L., and Carvalho, M.G., "A Recipe for Image Characterization of Fractal-Like Aggregates," J. Aerosol Sci. 30(10):1379-1389, 1999.
[35.] Pickett, L.M. and Siebers, D.L., "Soot in Diesel Fuel Jets: Effects of Ambient Temperature, Ambient Density, and Injection Pressure," Combust. Flame 138(1-2):114-135, 2004.
[36.] Dobbins, R.A. and Megaridis, C.M., "Absorption and Scattering of Light by Polydisperse Aggregates,"Appl. Opt. 30:4747-4754, 1991.
[37.] Sorensen, C.M., "Light Scattering by Fractal Aggregates: A Review," Aerosol Sci. Technol. 35:648-687, 2001.
[38.] Williams, T.C., Shaddix, C.R., Jensen, K.A., and Suo-Anttila, J.M., "Measurement of the Dimensionless Extinction Coefficient of Soot within Laminar Diffusion Flames," Int. J. Heat Mass Trans 50:1616-1630, 2007.
[39.] Harauz G. and Fong-Lochovsky, A.. "Automatic Selection of Macromolecules from Electron Micrographs by Component Labelling and Symbolic Processing," Ultramicroscopy 31(4):333-344, 1989.
[40.] Lata, K.R., Penczek, P., and Frank, J. "Automatic Particle Picking from Electron Micrographs," Ultramicroscopy 58(3-4):381-391, 1995.
[41.] Grishin, I., Thomson K., Migliorini, F., and Sloan, J.J., "Application of the Hough transform for the Automatic Determination of Soot Aggregate Morphology,"Appl. Opt. 51(5):610-620, 2012.
[42.] Hough, P.V.C., "Methods and Means for Recognizing Complex Patterns," U.S. Patent 3069654A, 1962.
[43.] Duda, R.O. and Hart, P.E., "Use of the Hough Transform to Detect Lines and Curves in Pictures," Communications of the ACM 15(1): 11-15, 1972.
[44.] Canny, J., "A Computational Approach to Edge Detection," IEEE Trans. Pattern Analysis Machine Intelligence 8(6):679-698, 1986.
[45.] De, S., Biswas, N., Sanyal, A., Ray P. et al., "Detecting Surface Circular Objects from Low Contrast Noisy Images: Applications in Microscope Image Enhancement," World Academy of Science, Engineering and Technology 6:2012-07-26, 2012.
[46.] Engine Combustion Network Data Archive: http://www.sandia.gov/ecn/dieselspraycombustion
[47.] Bruneaux, G., "Mixing Process in High Pressure Diesel Jets by Normalized Laser Induced Exciplex Fluorescence Part I: Free Jet," SAE Technical Paper 2005-01-2100, 2005, doi:10.4271/2005-01-2100.
[48.] Cenker, E., Bruneaux, G., Pickett, L., and Schulz, C., "Study of Soot Formation and Oxidation in the Engine Combustion Network (ECN), Spray A: Effects of Ambient Temperature and Oxygen Concentration," SAE Int. J. Engines 6(1):352-365, 2013, doi:10.4271/2013-01-0901.
[49.] Cenker, E., Kondo, K., Bruneaux, G., Dreier, T. et al., "Assessment of Soot Particle-Size Imaging with LII at Diesel Engine Conditions," submitted to Applied Physics B, 2015.
[50.] Lequien, G., Andersson, O., Tunestal, P., and Lewander, M., "A Correlation Analysis of the Roles of Soot Formation and Oxidation in a Heavy-Duty Diesel Engine," SAE Technical Paper 2013-01-2535, 2013, doi:10.4271/2013-01-2535.
[51.] Rissler, J., Messing, M.E., Malik, A.I., Nilsson, P.T. et al., "Effective Density Characterization of Soot Agglomerates from Various Sources and Comparison to Aggregation Theory," Aerosol Sci Tech 47(7):792-805, 2013.
[52.] Andersson, O., "Handbook of Combustion: Diesel Combustion Chapter," 3:415-440, Weinheim: Wiley-VCH Verlag GmbH & Co. 2010.
[53.] Chan, H.-P., Vyborny, C.J., MacMahon, H., Metz, C.E. et al., "Digital Mammography: ROC Studies of the Effects of Pixel Size and Unsharp-Mask Filtering on the Detection of Subtle Microcalcifications," Invest Radiol 22:581-589, 1987.
[54.] Gonzalez, R.C. and Woods, R.E., Digital Image Processing, 3rd edition, Prentice Hall, 2008.
% Automatic primary particle finder v1.2 (works on Matlab 2012a or higher + Image Processing Toolbox)
% The code implements pre-processing (Median Filter and unsharp masking), Canny edge detection, and
% Circular Hough Transform.
% Original code written by Qing Nian Chan on 18 Sep 2014
% Modified by Sanghoon Kook for diesel soot applications on 27 Sep 2014
% Last update on 19 Dec 2014 by Sanghoon Kook
close all;clear all;
TEMscale = 1; % e.g. 200 nm per 200 pixels in the scale bar
maxImgCount = 255; % Maximum image count for 8-bit image
SelfSubt = 0.8; % Self-subtraction level
mf = 1; % Median filter [x x] if needed
alpha = 0.1; % Shape of the negative Laplacian "unsharp" filter 0~1
rmax = 30; % Maximum radius in pixel
rmin = 4; % Minimun radius in pixel
sens_val = 0.75; % the sensitivity (0~1) for the circular Hough transform
ImgFile = ['TEMimage.tif']; % soot TEM image
% If dpAutomatedDetection is called up as a function...
%[dpdist] = dpAutomatedDetection(TEMscale,maxImgCount,SelfSubt,mf,alpha,rmin,rmax,sens_val,ImgFile);
%function[dpdist] = dpAutomatedDetection(TEMscale,maxImgCount,SelfSubt,mf,alpha,rmin,rmax,sens_val,ImgFile)
OriginalImg = II1;
%% - step 1: invert
if size(OriginalImg,1) > 900
II1(950:size(II1,1), 1:250) = 0;% ignore scale bar in the TEM image x 1-250 pixel and y 950-max pixel end
II1_bg=SelfSubt*II1; % Self-subtration from the original image
figure();imshow(II1, );title('Step 1: Inversion and self-subtraction');
% - step 2: median filter to remove noise
II1_mf=medfilt2(II1, [mf mf]);
figure();imshow(II1_mf, );title('Step 2: Median filter');
% - step 3: Unsharp filter
f = fspecial('unsharp', alpha);
II1_lt = imfilter(II1_mf, f);
figure();imshow(II1_lt, );title('Step 3: Unsharp filter');
%% Canny edge detection
BWCED = edge(II1_lt,'canny');
figure();imshow(BWCED);title('Step 4: Canny edge detection');
%% Find circles within soot aggregates
[centersCED, radiiCED, metricCED] = imfindcircles(BWCED,[rmin rmax], 'objectpolarity', 'bright', 'sensitivity', sens_val, 'method', 'twostage'); % - draw circles
figure();imshow(OriginalImg,);hold;h = viscircles(centersCED, radiiCED, 'EdgeColor','r');
title('Step 5: Parimary particles overlaid on the original TEM image');
%% - check the circle finder by overlaying the CHT boundaries on the original image
R = imfuse(BWCED, OriginalImg,'blend');
figure();imshow(R,,'InitialMagnification',500);hold;h = viscircles(centersCED, radiiCED, 'EdgeColor','r');
title('Step 6: Primary particles overlaid on the Canny edges and the original TEM image');
dpdist = radiiCED*TEMscale*2;
save([ImgFile '_dp.mat'], 'dpdist_CED', 'centersCED', 'metricCED'); % Save the results
%end % for the dpAutomatedDetection function
Sanghoon Kook, Renlin Zhang, and Qing Nian Chan
Univ. of New South Wales
Tetsuya Aizawa and Katsufumi Kondo
Lyle M. Pickett
Sandia National Laboratories
Emre Cenker and Gilles Bruneaux
Oivind Andersson, Joakim Pagels, and Erik Z. Nordin
Table 1. Meiji University CVC operating conditions Ambient gas conditions Density 9.5kg/[m.sub.3] Temperature 940 K Pressure 2.5 MPa [O.sub.2] 21% Injection conditions Type Common-rail Nozzle Single-hole, mini-sac Nozzle diameter 140 [micro]m (nominal) Injection pressure 80 MPa Injection duration 2.5 ms Fuel JIS#2 Mass per injection 10.3 mg Sampling location from nozzle 70 mm (peak soot) Table 2. Sandia CVC operating conditions Ambient gas conditions Density 22.8 kg/[m.sub.3] Temperature 1000 K Pressure 6.7 MPa [O.sub.2] 15% Injector conditions Type Bosch common-rail second- generation injector Nozzle Single-hole, KS1.5/0.86, mini-sac Nozzle diameter 90 [micro]m (nominal) Injection pressure 150 MPa Injection duration 7 ms Fuel No. 2 Diesel Sampling location from nozzle 50 mm (peak soot) Table 3. IFPEN CVC operating conditions ([O.sub.2] Variants of ECN Spray A target conditions ) Ambient gas conditions Density 22.8 kg/[m.sup.3] Temperature 900 K Pressure 6.0 MPa [o.sub.2] 21% Injector conditions Type Bosch common-rail second- generation injector Nozzle Single-hole, KS1.5/0.86, mini-sac Nozzle diameter 90 [micro]m (nominal) Injection pressure 150 MPa Injection duration 6 ms Fuel N-dodecane Sampling location from nozzle 45 mm (peak soot) Table 4. UNSW single-cylinder optical light-duty diesel engine specifications and operating conditions Number of cylinders 1 Displacement volume per 498 [cm.sub.3] cylinder Bore / Stroke 83 mm / 92 mm Compression ratio 15.2 Swirl ratio 1.4 Coolant temperature 363 K Intake air temperature 303 K Intake [O.sub.2] 21% Engine speed 1200 rpm Fuel Ultra-low-sulphur diesel Cetane number 51 Injection system Bosch common-rail second-generation injector Nozzle type Single-hole, K1.5/0.86 Hole diameter 134 [micro]m (nominal) Injection pressure 70 MPa Injection duration 2.34 ms (actual) Mass per injection 9 mg Injection timing -7[degrees]CAaTDC Table 5. Lund University six-cylinder heavy-duty diesel engine specifications and operating conditions Number of cylinders 6 Displacement volume per 2130 [cm.sub.3] cylinder Bore / Stroke 131 mm /158 mm Compression ratio 16 Exhaust gas recirculation 43.3% Intake [O.sub.2] 10.4% Engine speed 1200 rpm Fuel Ultra-low-sulphur diesel (MK1) Cetane number 51 Injection system Delphi E3 Unit Injector Nozzle type 6 holes Hole diameter 219 [micro]m (nominal) Injection pressure 200 MPa maximum Equivalence ratio 0.645 IMEP 1 MPa Combustion phasing (CA50) 3[degrees]CA aTDC Table 6. Image-processing parameters and set values for the Meiji CVC and the UNSW diesel engine samples CVC Engine TEM scale fnm/pixell 0.464 1 Maximum image count 5568 255 1. Self-subtraction [0-11 1.2 0.8 Median filtering [pixel by pixel] 1 1 2. Negative Laplacian (Unsharp) filter shape [0-11 0.1 0.1 3. Maximum radius of primary particles [pixel] 45 30 Maximum diameter of primary particles [nm] 41.74 60 4. Minimum radius of primary particles [pixel] 9 4 Vlinimum diameter of primarv particles [nm] 8.35 8 5. Circular Houqh Transform (CHT) sensitivity [0-1] 0.79 0.75 Table 7. Image-processing parameters and set values for the Sandia and IFPEN CVC, and Lund University diesel engine (exhaust) samples Sandia IFPen Lund TEM scale [nm/pixel] 1 0.464 0.464 Maximum image count 255 18k 19k 1. Self-subtraction [0-11 0.5 1.2 1.2 Median filterinq [pixel bv pixel] 1 1 1 2. Neqative Laplacian (Unsharp) filter shape [0-1] 0.1 0.1 0.1 3. Maximum radius of primary particles [pixel] 30 45 45 Maximum diameter of primary particles [nm] 60 41.74 41.74 4. Minimum radius of primary particles [pixel] 4 9 9 Minimum diameter of primary particles [nm] 8 8.35 8.35 5. Circular Houqh Transform (CHT) sensitivity [0-1] 0.75 0.79 0.79
|Printer friendly Cite/link Email Feedback|
|Author:||Kook, Sanghoon; Zhang, Renlin; Chan, Qing Nian; Aizawa, Tetsuya; Kondo, Katsufumi; Pickett, Lyle M.;|
|Publication:||SAE International Journal of Engines|
|Article Type:||Technical report|
|Date:||Apr 1, 2016|
|Previous Article:||Experimental Investigations of a DISI Engine in Transient Operation with Regard to Particle and Gaseous Engine-out Emissions.|
|Next Article:||Investigation of Soot Oxidation Carried out on Membrane Filters Composed of SiC Nanoparticles.|