(MATLAB Analysis tools for TEM images of Soot)
This codebase contains MATLAB code for several methods of characterizing soot aggregates in TEM images. This includes methods for evaluating the aggregate projected area, perimeter, and primary particle diameter. Methods include Otsu thresholding, the pair correlation method (PCM), Hough circle transform (following Kook et al. (2015)), the Euclidean distance mapping, scale-based analysis (EDM-SBS) method of Bescond et al. (2014), and tools to aid in manual analysis. This code is designed to replace a previous, deprecated code.
Testing of this codebase makes use of the main_*
functions in the upper directory, which are described in 1. MAIN SCRIPTS (main_*) below. Specifically, main_kmeans
and main_auto
test the fully automated methods, while main_0
allows for testing of the more manual methods (which require substantial user input).
The program is primarily composed of two analysis packages, which will be discussed later in the README:
-
+agg - which performs aggregate-level segmentation to output a binary image, and
-
+pp - which determines primarily particle information, often from the binary image generated by the methods in the agg package noted above.
This code also includes a set of utility functions in the +tools package and Python code necessary to implement a convolutional neural network used for segmentation in the carboseg folder. The corresponds ONNX file is temporarily available here: https://github.com/tsipkens/FPN-resnet50-imagenet.
A. Dependencies: What do I need to run this codebase?
B. Getting started: Walking through a sample code
C. Components: Details on the structure of this codebase
- 1. Main scripts (main_*): Testing the codebase
- 2. Aggregate segmentation package (+agg)
- carboseg: Using a Python-based neural network
- 3. Primary particle analysis package (+pp)
- 4. Additional tools package (+tools)
D. Backmatter: Contributors, acknowledgements, how to cite, and references
This software was tested using MATLAB 2022a (though most functions have been validated against older versions). The current code depends on the following MATLAB toolboxes:
- the curve fitting toolbox,
- the financial toolbox,
- the image toolbox,
- the optimization toolbox, and
- the video and image blockset (in the computer vision toolbox, which is used for OCR).
If not already installed, these packages can be added to MATLAB via the instructions available on MATLAB's website.
This program also has an optional submodule that adds support for perceptually uniform colormaps compiled from various sources and available at https://github.com/tsipkens/cmap. It can be added via git using
git submodule add -b master https://github.com/tsipkens/cmap
git submodule init
The submodule is not necessary for any of the scripts or methods included with this code.
Additional dependencies are required for use of the carboseg or convolutional neural network component of this program, including a copy of Python. See the appropriate section below for more information.
This code has been tested on Windows 10+. Linux users may have some issues with loading images when supplying a string to tools.load_imgs(...)
(for example, because of differences in file separator). macOS users may encounter similar issues, including in writing aggregate structures to JSON files and in certain plotting functions.
The overall structure of scripts associated with this code can be broken down into three steps.
The first step in the image analysis process is to import images. Images can be handled in one of two ways.
The first is as a MATLAB structure, with one entry in the structure for each image loaded. To generate this structure, one can use the tools.load_imgs(...)
function, either giving a folder name as an input or no argument (to use a file explorer). To use the test images:
Imgs = tools.load_imgs('images'); % load the images
The output structure contains the image file name and directory; the image itself, with and without the footer removed (or cropped
); and the pixel size, read from the image footer. Alternatively, one can use a file explorer to select the folders by excluding any arguments:
Imgs = tools.load_imgs; % load the images
Second, the image can also be handled using a cell array of cropped images and pixel sizes. These are secondary outputs to the tools.load_imgs(...)
function:
[~, imgs, pixsizes] = tools.load_imgs('images'); % load the images
The images and pixel sizes can equivalently be extracted from the Imgs
structure using:
imgs = {Imgs.cropped}; % copy variables locally
pixsizes = [Imgs.pixsize]; % pixel size for each image
fname = {Imgs.fname};
We note that detecting the footer and pixel size (or scale) uses the detect_footer_scale(...)
subfunction of the tools.load_imgs(...)
method. This function attempts to interpret the image footer and/or scale using image analysis tools. The method is known to work with TEM images taken at the University of British Columbia, where it applies optical character recognition to determine the pixel size from the footer text (stored in the Imgs.pixsize
field and/or output directly as pixsizes
) and crops the footer away. Black text and a scale bar overlaid on the image may also be readable, but present more challenges such that determining the pixel size is not always successful. In cases where this latter approach is used, the code will attempt to fill the overlaid scale elements with background noise, to improve subsequent analyses. Should all of this fail, the user can also call the tools.ui_scale_bar(imgs)
function to use a UI to estimate the pixel size for the cell of raw images specified by imgs
. Type
help tools.ui_scale_bar;
on the MATLAB command line for more information on that function.
NOTE: At the moment, this approach will load all of the images into memory. For computers with less memory, this could restrict the number of images that can be analyzed at one time. Batches of 250 images have been run successfully on a computer with 16 GB of RAM (MATLAB limits memory usage to below this value). A simple work around is to load the images in groups, rather than all at once, and then add an outer loop to proceed through the different groups. The
tools.load_imgs(fd, n)
function is equipped to handle this, passing the second, optional argumentn
as the integer range of images to load into memory. For example,n = 1:3
will load the first three images from the set specified byfd
.
The next step is to evaluate binary masks that separate the image into pixels that are part of the background and pixels that are part of aggregates. This is done using the functions in the +agg package. For example, a k-means classifier can be used by calling:
imgs_binary = agg.seg_kmeans(imgs, pixsizes);
% segment aggregates
The result, imgs_binary
, is either a single binary image (if one image is given as an input) or a cell array of image binaries (if multiple images are given as an input), with 1
if a pixel is considered part of the aggregate and 0
if it is not. Other approaches are also available and are discussed in Section 2 below.
Having segmented the image, aggregate characteristics can be determined by passing this binary image to the tools.analyze_binary(...)
function, e.g.:
Aggs = agg.analyze_binary(...
imgs_binary, pixsizes, imgs, fname);
% determine aggregate properties
The output data itself, Aggs
, is a MATLAB structure with one entry per aggregate, containing properties like the location of the aggregate, perimeter, and projected-area. This data can then be exported to a JSON file using :
tools.write_json(Aggs, fname);
or an Excel spreadsheet using:
tools.write_excel(Aggs, fname);
to be analyzed in other software and languages, where fname
is the filename of the file to write to. The Aggs
structure can also be visualized as an overlay on top of the TEM image using
figure(1);
tools.imshow_agg(Aggs);
The resultant image highlights pixels that are part of the aggregate, and plots a circle that corresponds to the radius of gyration. This output is similar to that shown the image in the header of this README.
Primary particle size information can finally be determined using functions in the +pp package. The original pair correlation method (PCM), for example, can be applied by using the Aggs
output from the agg.analyze_binary(...)
function as
Aggs = pp.pcm(Aggs); % apply pair correlation method
The output is an updated Aggs
structure that also contains an estimate of the primary particle size for each aggregate. Having done this, the primary particle size can be visualized along with the radius of gyration noted above by passing the updated Aggs
structure to the tools.plot_aggregates(...)
function:
figure(2);
tools.imshow_agg(Aggs);
The inner circle in this plot now indicates the primary particle size from PCM, the larger circle the radius of gyration from k-means, and the number indicating the index used by the program to identify each aggregate. Images produced using this type of procedure will feature heavily in the remainder of this README.
Code associated with this example is given in the script main_b.m. The output should correspond to the figure shown below.
This section provides details of the overall structure of this codebase and the methods available for the characterization of soot aggregates in TEM images.
The main scripts demonstrate further use of the code for specific scenarios and are provided to test the codebase. Three such scripts are included:
main_b is the script associated with the demonstration in B. Getting started above.
main_kmeans is designed to specifically investigate the k-means approach to aggregate-level segmentation. By default, this is done on the sample images provided in the images folder included with this distribution. This script will also read in some binaries (provided in the images/slider folder) produced by the slider method beforehand for comparison. Finally, the primary particle size is computed for the k-means binaries and is plotted for the user. This script supports a European Aerosol Conference submission (Sipkens et al., 2020).
main_auto runs through most of the fully automated techniques provided with this program (e.g., k-means, Otsu, PCM, etc.), applying them to the images provided in the images folder. The binary images produced by each method are overlaid on the original images for inspection by the user. Outputs are similar to those presented in Section 2 below.
main_0 script presents use of the agg.seg(...)
general segmentation function, described below, as well as how to read in one's own images. At the end of the script, data is written to an Excel file for examination external to MATLAB. Note that this requires that the user's images be formatted appropriately, either with a horizontal footer below the TEM image that has a white background or a cropped TEM image, in which case the user will have to provide pixel size information. Sample images on which the method have been tested are again found in the images folder.
This package contains an expandable library of functions aimed at performing semantic segmentation of the TEM images into aggregate and background regions. Functions are accessed by appending agg.
to the function name, e.g., agg.seg_kmeans(...)
to call the k-means segmentation procedure.
These images are included with this distribution in the images/
folder. These images represent soot collected from a lab-scale flare (Trivanovic et al., 2020) and a diesel engine (Kheirkhah et al., 2020).
The main functions implementing aggregate-level semantic segmentation have filenames following the template agg.seg_*
. In each case, the output primarily consists of binary images of the same size as the original image but where pixels take on logical values: 1
for pixels identified as part of the aggregate 0
for pixels identified as part of the background.
The functions often take similar inputs and provide a binary mask output. As a general template,
img_binary = agg.seg_slider_orig(imgs, ...)
where the imgs
input is one of:
- a single image (after any footer or additional information have been removed);
- an
Imgs
structure, containing the image information as fields; or - a cell array of images (again with the footer removed).
and the img_binary
output is a cell of binary masks. Several methods also take pixsize
, which denotes the size of each pixel in the image. If an Imgs
structure if provided, this information is expected to be contained in this structure and any pixsize
input is ignored. Other arguments depend on the function (e.g., optional parameters for the rolling ball transform). For input arguments relevant to any given method, please refer to function headers and/or definitions.
Several of the available methods are summarized briefly below.
This function applies a k-means segmentation approach following Sipkens and Rogak (2021) and using three feature layers, which include: (i) a denoised version of the image; (ii) a measure of texture in the image; and (iii) an Otsu-like classified image, with the threshold adjusted upwards. This method works well for a large range of images, e.g.,
Though, the technique still occasionally fails, particularly if the function does not adequately remove the background. The method also has some notable limitations when images are (i) zoomed in on a single aggregate while (ii) also slightly overexposed. The k-means method is associated with configuration files (cf., +agg/conifg/), which include different versions and allow for tweaking of the options associated with the method. See the seg_kmeans(...)
function header or type help agg.seg_kmeans
for more information, including configuration versions and function arguments.
This seg_carboseg(...)
function employs Python to implement a convolutional neural network (CNN) for segmentation as described by Sipkens et al. Details and code for the training of the network are available in a parallel repository at https://github.com/maxfrei750/CarbonBlackSegmentation, with primary contributions by Max Frei (@maxfrei750). The implementation here makes use of the ONNX file output (to be downloaded here) from that procedure and employs the Python ONNX runtime for execution. Use of this function requires the necessary Python environment as a pre-requisite.
We also note that, as of this writing, MATLAB does not support the necessary layers to import the ONNX as a native MATLAB object.
Details on the setup and use of this component are given in the README in the carboseg subfolder. Two options exist for interacting with the Python code: (1) initializing an instance of Python directly in MATLAB or (2) a multi-step procedure of exporting and reimporting images.
These automated methods apply Otsu thresholding followed by a rolling ball transformation. Two versions of this function are included.
A. The agg.seg_otsu_orig(...)
function remains more true to the original code of Dastanpour et al. (2016). For the sample images, this often results in fragmented segmentations, e.g.,
As the technique may be insufficient on its own, this implementation can be complimented with agg.seg_slider(...)
, described below, to fill in the gaps between the aggregates and add missing aggregates.
B. Stemming from the deficiencies of the above function, the agg.seg_otsu(...)
function updates the above implementation by (i) not immediately removing boundary aggregates, (ii) adding a background subtraction step using the agg.bg_subtract(...)
function, and (iii) adding a bilateral denoising step. This results in the following segmentations.
This latter function generally performs better, though the results are still significantly fragmented. The technique generally underperforms relative to the previously mentioned k-means method but acts as a good way to initialize more manual techniques.
The agg.seg_slider_orig(...)
method is a largely manual technique originally developed by Ramin Dastanpour (Dastanpour et al., 2016)). The function enacts a GUI-based method with a slider for adaptive, semi-automatic thresholding of the image (adaptive in that small sections of the image can be cropped and assigned individually-selected thresholds). It is worth noting that the mostly manual nature of this approach will resulting in variability and subjectiveness between users but that the human input often greatly improves the quality of the segmentations. See the seg_slider_orig(...)
function header or type help agg.seg_slider_orig
, including usage.
We note that this code saw important bug updates since the original code by Dastanpour et al. (2016). This includes fixing how the original code would repeatedly apply a Gaussian filter every time the user interacted with the slider in the GUI (which may cause some backward compatibility issues), a reduction in the use of global variables, memory savings, and other performance improvements.
The core of the seg_slider(...)
method is the same as the GUI-based slider method described above but sees an overhaul of the user interface. This implementation makes use of MATLAB's app builder, requiring newer MATLAB versions to use.
The agg.seg(...)
function is a general, multipurpose wrapper function that attempts several methods listed here in sequence, prompting the user after each attempt. Specifically, the method attempts (i) the k-means classifier, (ii) followed by the Otsu classifier, and finally (iii) reverts to using the slider method. This is repeated until the user has classified all of the images that were passed to the function.
All of the above methods produce a common output: a binary image. The agg.analyze_binary(...)
function is now used to convert these binaries to aggregate characteristics, such as area in pixels, radius of gyration, area-equivalent diameter, aspect ratio, among other quantities. The function itself takes a binary image, the original image, and the pixel size as inputs, as follows.
Aggs = agg.analyze_binary(imgs_binary, pixsize, imgs, fname);
The output is a MATLAB structured array, Aggs
, containing information about the aggregate. The array has one entry for each aggregate found in the image, which is itself defined as any independent groupings of pixels. The fname
argument is optional and adds this tag to the information in the output Aggs
structure.
Multiple of these methods make use of the rolling ball transformation, applied using the agg.rolling_ball(...)
function included with this package. This transform fills in gaps inside aggregates, acting as a kind of noise filter. This is accomplished by way iterative morphological opening and closing.
The +pp package contains multiple methods for determining the primary particle size of the aggregates of interest. Often this requires a binary mask of the image that can be generated using the +agg package methods. After applying most of the methods, the primary particle size will be stored in (i) the Aggs.dp
field and (ii) another Aggs
field with additional information specifying the method used (e.g., Aggs.dp_pcm1
contains a PCM-derived primary particle diameter). For the former, whichever method was last used will overwrite the Aggs.dp
field, which then acts as a default value that is used by other functions (by the tools.imshow_agg(...)
function).
The pp.pcm
function contains code for the University of British Columbia's pair correlation method (PCM) method, originally developed by Ramin Dastanpour (Dastanpour et al. (2016)). This package contains a significant improvements to code readability, memory use, and length relative to the previous code. The underlying method is largely unchanged, correlating the relationship between pixels to the primary particle size for a given aggregate. A single average primary particle diameter is given for each aggregate. Dastanpour et al. provided two different types of pair correlation function (PCF), corresponding to Aggs.dp_pcm1
, previously denoted as simple and Aggs.dp_pcm2
, previously denoted as general. Testing has generally suggested that the simple method perform better, and this value is assigned to Aggs.dp
in the output from the PCM method.
The Euclidean distance mapping, scale-based analysis (EDM-SBS) of Bescond et al. (2014) is implement in the pp.edm_sbs(...)
function. This is an adaptation of the original code for use with MATLAB and using the binaries above in the place of output from imageJ. As such, some minor differences in output should be expected (which are challenging to compare, as the ImageJ output does not have a direct analog here). The method remains true to how it is described in Bescond et al. and ports some components from the original Scilab code (version 3, available here).
Among the changes to the original EDM-SBS code, this implementation also applies the EDM-SBS method to individual aggregates. While this allows for a better comparison to the other methods here, the method was originally intended to evaluate the primary particle size distribution across a range of aggregates, with uncertaintainty ramifications. However, the linear nature of the curves generated by the method means that the overall EDM-SBS curve is simply approximated by the superposition of all of the aggregates. This is also output by the present code.
Two pp.hough_kook**(...)
functions are included with this program, which fit circles to features in the image using the Hough transform and the pre-processing steps described by Kook et al. (2015).
The first function, pp.hough_kook*(...)
, contains a copy of the code provided by Kook et al. (2015), with minor modifications to match the input/output of some of the other packages — namely to take a single image, img
, and a pixel size, pixsize
— and to output a Pp
structure, which contains information for each circle. Note that the original function acts on images without trying to assign primary particles to an aggregate, something resolved in the second function below. This causes some compatibility issues in terms of comparing the output from this function to the other methods contained in this program.
The pp.hough_kook2(...)
function contains a modified version of the method proposed by Kook et al. (2015) that excludes circles in the background and assigns primary particles to aggregates. This is done rather simply: by checking if the center of the circles from the preceding procedure lie within the binary for a given aggregate. A sample output is as follows.
Here, red circles are identified as part of an aggregate, while black circles are excluded in the output. Note that it is apparent from some of the images that gradients in the background influenced the results. Though, it can also be noted that the circles within the aggregates remain reasonable (to the extent that the overall method is reasonable).
The pp.manual(...)
function can be used to manual draw circles around the soot primary particles. The code was developed at the University of British Columbia and represents a heavily modified version of the code associated with Dastanpour and Rogak (2014). The current method uses two lines overlaid on each primary particle to select the length and width of the particle. This is converted to various quantities, such as the center of each primary particle and the overall mean primary particle diameter.
This package contains a series of functions that help in visualizing or analyzing the aggregates that transcend multiple methods or functions. We discuss a few examples here and refer the reader to individual functions for more information.
** These functions aid in visualizing the resultant images. For example,
tools.imshow_binary(img, img_binary);
will plot the image, given in img
, and overlay labels for the aggregates in a corresponding binary image, given in img_binary
. Appending an opts
structure to the function allows for the control of the appearance of the overlay, including the label alpha and colour. For example,
opts.cmap = [0.92,0.16,0.49];
tools.imshow_binary(img, img_binary, opts);
will plot the overlays in a red, while
opts.cmap = [0.99,0.86,0.37];
tools.imshow_binary(img, img_binary, opts);
will plot the overlays in a yellow.
The related tools.imshow_agg(...)
function will plot the binaries, as above, and add aggregate-level information to the plot, including: (i) the aggregate number; (ii) the radius of gyration about the center of the aggregate; and (iii) the average primary particle diameter, if this information is available (the function will use the Aggs.dp
field, which will contain primary particle information from the most recently applied method).
This set of functions writes data to files, with the precise format depending on the function.
The tools.write_excel(...)
and tools.write_json(...)
functions write aggregate-level information to Excel and JSON formats respectively. The precise output will depend on what information is contained in the Aggs
structure given to the method. For example, Aggs
structures with primary particle information will have that information output to the Excel or JSON files.
The tools.write_images(...)
function, in contrast, take a cell of images (or a single image) and writes them to a series of files, specified by the fnames
argument, in the folder specified by the folder
argument. For example, this can be used to write the binary images presented at the beginning of this README to a temporary folder using
tools.imwrite(imgs_binary, fname, 'temp');
This function is also useful when paired with the tools.imshow_agg(...)
function to write images with binary overlays and radius of gyration information (such as those presented throughout this README). For a series of images that have been processed to produce an Aggs
structure, this can be accomplished using:
imgs_agg{length(imgs)} = []; % initialize cell
% Loop through images and get overlay for each image.
% The second argument of tools.imshow_agg gets the
% colordata from the frame in the overlay image.
for ii=1:length(imgs)
[~, imgs_agg{ii}] = tools.imshow_agg(Aggs, ii, 1, opts);
end
% Write the overlay images to a temporary folder
tools.write_images(imgs_agg, fname, 'temp');
As the image size will depend on the physical size of the figure on the screen, it may be useful to first run
f1 = figure(1); f1.WindowState = 'maximized';
to maximize the figure window, to generate higher quality output.
This set of functions is intended to generate formatted plots of post-processed results. For example, tools.viz_darho
generates a scatter plot of the area-equivalent diameter versus effective density estimated for each aggregate.
This software is relaesed under an MIT license (see the corresponding license file for details).
This code was primarily compiled by Timothy A. Sipkens while at the University of British Columbia (UBC), who can be contacted at [email protected]. This work was supported by Dr. Steven Rogak ([email protected]). Other direct contributors include Hamed Nikookar and Darwin Zhu.
Pieces of this code were adapted from various sources and features snippets written by several individuals at UBC, including Ramin Dastanpour, Una Trivanovic, Alberto Baldelli, Yiling Kang, Yeshun (Samuel) Ma, and Steven Rogak, among others.
This program contains very significantly modified versions of the code distributed with Dastanpour et al. (2016). The most recent version of the Dastanpour et al. code prior to this overhaul is available at https://github.com/unatriva/UBC-PCM (which itself presents a minor update from the original). That code forms the basis for some of the methods underlying the manual processing and the PCM method used in this code, as noted in the README above. However, significant optimizations have improved code legibility, performance, and maintainability (e.g., the code no longer uses global variables).
Also included with this program is the MATLAB code of Kook et al. (2015), modified to accommodate the expected inputs and outputs common to the other functions.
This code also contain an adaptation of EDM-SBS method of Bescond et al. (2014). We thank the authors, in particular Jérôme Yon, for their help in understanding their original Scilab code and ImageJ plugin. Modifications to allow the method to work directly on binary images (rather than a custom output from ImageJ) and to integrate the method into the MATLAB environment may present some minor compatibility issues, but allows use of the aggregate segmentation methods given in the agg package.
The carboseg method follows from collaborative work with Max Frei and is associated with Sipkens et al. (2021).
Finally, the progress bar in the function tools.textbar
, which is used to indicate progress on some of the primary particle sizing techniques, is a modified version of that written by Samuel Grauer.
On use of this code or the k-means segmentation procedure described above, please cite:
Users of the pair correlation method (PCM), the Euclidean distance mapping-surface based scale (EBD-SBS), and Hough transform (following Kook et al.) codes should acknowledge the corresponding studies under the acknowledgements above. Please also consider citing this repository directly.
When using the CNN method (e.g., carboseg) please cite
and see the CarbonBlackSegmentation repository for information on training.