Minerals Engineering 193 (2023) 108019

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Online particle size analysis on conveyor belts with dense convolutional

neural networks
Yihao Fu a, Chris Aldrich a, b, *
a
Western Australian School of Mines: Minerals, Energy and Chemical Engineering, Curtin University, GPO Box U1987, WA 6845, Australia
b
Department of Process Engineering, Stellenbosch University, Private Bag X1, Matieland 7602, Stellenbosch, South Africa

A R T I C L E I N F O A B S T R A C T

Keywords: Particle size distributions in ore feed systems, as well as the identification of particles in these feed systems can
Particle size analysis provide important information in the advanced control of unit operations in mineral processing, such as crushing
Bulk solids characterization and grinding circuits. Image analysis has long been considered a promising approach to achieve this, as it is an
Image analysis
inexpensive, unobtrusive means of acquiring information rich measurements. It typically requires segmentation
U-net
Semantic labeling
of images in order to identify individual particles. This is a challenging task to accomplish reliably, as variable
Watershed lighting, fines adhering to larger particles or contiguous particles, as well as variable particle sizes and shapes can
Superpixels all compromise the accuracy of traditional algorithms. Image segmentation with deep learning methods have
Grinding circuits recently been investigated to surmount these difficulties. In this investigation, U-net and U-net with superpixel
Deep learning preprocessing with simple linear iterative clustering (SLIC) are proposed and compared with a traditional
Simple linear iterative clustering watershed algorithm. The U-net approaches were markedly more reliable than the watershed algorithm. In
addition, preprocessing of the images with SLIC resulted in further improvement of the results.

1. Introduction conditions, dust, etc.

Image analysis is an attractive technology with regard to online
The identification of individual particles in the feed and product particle characterization, since it is comparatively inexpensive, nonin­
streams of mineral processing operations is key to the efficient operation vasive and provides information intensive measurement. The use of
of many diverse process circuits, such as in the coal preparation industry machine learning in image segmentation of particle systems is a recent
(Zhang et al., 2012; 2013; 2020; Bai et al., 2021; Qiu et al., 2021), development.
comminution circuits (Guyot et al., 2004; Kaartinen and Tolonen, 2008; In earlier machine learning applications, particle size estimation was
Leroy et al., 2011), oil sands (Mukherjee et al., 2009), ore sorting mostly treated as pattern recognition problems, where images are
(Manana et al., 1985; Singh and Rao, 2005), preconcentration of ores treated as patterns that can be associated with the bulk characteristics of
(Sousa et al., 2020) and rock blasting (Kulatilake et al., 2010). the ore. These applications included determination of the presence or
Traditionally, manual screening methods are used to analyse particle absence of fines (Aldrich et al., 2009; Amankwah and Aldrich, 2011;
size distributions. Apart from the cost involved, this information is not Jemwa and Aldrich, 2012), a qualitative assessment of the average ore
generally useful in online control of equipment. As a consequence, on­ size or the type of ore (Kistner et al., 2013; Min et al., 2012; Jemwa and
line estimation of particle size distributions and identification of ores Aldrich, 2010, 2012; Chatterjee et al., 2010, 2013; Aldrich et al., 2010;
have attracted considerable interest over the last few decades (Al- 2011; Patel et al., 2017; Olivier et al., 2019, 2020; Bamford et al., 2021;
Thyabat et al., 2007; Thurley, 2009; 2011; Hamzeloo et al., 2014; Leiva Liu et al., 2021a, 2021b). While these approaches can be effective, they
et al., 2021). This is a challenging problem, as particles in mineral typically require large sets of labelled images, which could be very
processing streams can have widely varying shapes, sizes, and textures expensive or impractical to obtain.
and can be subject to agglomeration, sliming, dusty environments and A more effective approach is to classify the pixels in the images,
stratification. Moreover, these systems may often have to cope with instead of the images themselves. In this regard, they are similar to any
widely varying ambient conditions with variable lighting, wet and dry of a large variety of traditional methods, similarly designed to achieve

* Corresponding author at: Western Australian School of Mines: Minerals, Energy and Chemical Engineering, Curtin University, GPO Box U1987, WA 6845,
Australia.
E-mail address: [email protected] (C. Aldrich).

https://doi.org/10.1016/j.mineng.2023.108019
Received 23 November 2021; Received in revised form 3 February 2023; Accepted 5 February 2023
Available online 14 February 2023
0892-6875/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Fig. 1. Image segmentation based on preprocessing with SLIC superpixelation, followed by pixel classification with a dense convolutional neural network (U-net).

segregation of the image, ranging from simple thresholding methods to or adding noise to the images to create a larger more diverse training
more complex edge or region-based approaches. The use of convolu­ data set. Following this, the convolutional neural network model is
tional neural networks for image segmentation was proposed by Long trained (C), which may also involve several stages of hyperparameter
et al. (2014), who were the first to train these networks end-to-end for optimization and validation of the model. In the final step, the opti­
pixel-wise classification. The key innovation was to supplement con­ mized, validated model is deployed to generate segmented images (D).
ventional contracting networks with successive layers, in which pooling
operators are replaced with up-sampling operators, as well as connect­
ing early layers in the encoder or contracting section of the network to 2.1. Superpixelation based on simple linear iterative clustering (SLIC)
corresponding layers in the decoding section of the network.
In the mineral processing industry, application of convolutional Superpixel techniques, first referred to as such by Ren and Malik
neural networks in image segmentation has also started to attract strong (2003), generate groups of pixels that are similar in appearance (Stutz
interest, such as in the characterization of cement fragmentation (Chen et al., 2017). This is an attempt to obtain regions that represent mean­
et al. 2018), rock fragmentation (Karimpouli and Tahmasebi 2019, ingful descriptions with orders of magnitude fewer data than is the case
Schenk et al., 2019; Li et al., 2021), ore particle size estimation on when all pixels in the image are used. This in turn reduces the number of
conveyor belts (Liu et al., 2020; Wang et al., 2021), X-ray computed primitives in the image, and therefore also the complexity of recognition
tomography of rock samples (Varfolomeev et al., 2019), identification of tasks, while still preserving important image properties (Yang et al.,
coal and gangue particles (Gao et al., 2020); size estimation of flotation 2020). Generally, superpixelation is based on measures of colour simi­
froth bubbles (Fu and Aldrich, 2020), online size distribution measure­ larity and regional shapes.
ment of iron pellets based on the use of a nested U-net (Wu et al., 2021). Several superpixel-based CNN models have recently been proposed,
Since these applications can generally be seen as image pixel clas­ where the model is trained using precomputed superpixel regions. For
sification problems, they circumvent the requirement for large sets of example, Qin et al. (2018) partitioned computed tomographic images of
labelled training images, as the images themselves would typically the liver into superpixel regions, and trained a CNN model to classify
contains many thousands or millions of pixels. The problem in this case them into three classes including the interior of the liver, liver boundary
is often that of reliably labelling images. While simple delineation of and non-liver background. Mu et al. (2020) generated saliency maps
particles with well-defined geometries that are sparsely distributed on based on superpixel similarity, before feeding the maps into a CNN
conveyor belts can easily be dealt with, relatively low-definition images model for segmentation. Another approach is to train the network to
of irregular densely packed particles of mixtures of ores with a large learn the feature embedding space from the superpixel regions, which
variation in size distributions could pose more of a challenge. effectively attempts to learn the similarities between neighboring
This difficulty can be alleviated by preprocessing of the images. superpixels, and then merge similar groups for the segmentation (Liu
Authors such as Tomasi and Manduchi (1998), Buades et al. (2005), et al., 2018).
Varfolomeev et al. (2019) and Dan et al. (2021) have shown that image The two most important approaches to generate superpixels are
preprocessing can improve segmentation performance significantly. based on graphs (e.g. Shi and Malik, 2000; Felzenszwalb and Hutten­
In this paper, it is likewise shown that preprocessing of images, fol­ locher, 2004) and clustering methods. In this investigation, simple linear
lowed by the use of a dense convolution neural networks, or U-net iterative clustering (SLIC) was used, since clustering is simpler and
structures, can go a long way to meeting these challenges. Moreover, it is naturally includes the proximity of pixels in images. SLIC is arguably the
also shown that the explicit identification of pixels on the boundaries of most popular method (Achanta et al., 2012) and with this approach,
particles as an additional class can improve image segmentation. each pixel is characterised by a vector consisting of five components.
In Section 2 of the paper, image segmentation with convolutional Three of these represent the colour and two represent the position of the
neural networks is discussed, after which the analytical methodology pixel. With greyscale images, three-dimensional vectors only are
used in this paper is introduced in Section 3. In Sections 4 and 5, the required.
approach is illustrated by means of two case studies and in Section 6, the More formally, given an image I with dimensions w × h, segmenta­
conclusions of the investigation are presented. tion with superpixel algorithms attempt to divide I into a set
of K superpixels, S = {S1 , S2 , ⋯SK }. Each superpixel Si consists of a
2. Methodology group of pixels that are similar on some set of features. More specifically,
the similarity between pixels is defined as
The overall methodology to identify particles in an image is based on
D = dc + w1 ds /w2 (1)
image preprocessing to generate a superpixelated image, which is then
used as input to a dense convolutional neural network designed to In eq (1), parameters w1 and w2 control the compactness of the re­
classify the pixels in the image as belonging to particles or image gion, while w2 is also a scaling factor. dc and ds are the colour and spatial
background, as indicated in Fig. 1. differences between the pixels, i.e. for pixels pi and pj , with colour co­
This methodology consists of four basic steps. In the first, training ordinates ri , gi and bi and rj , gj and bj respectively,
and test sets of images are preprocessed to generate superpixelated ( ) √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2 ( )2 ( )2̅
images (A). These images are subsequently labelled or provided with dc pi , pj = ri − rj + gi − gj + bi − bj (2)
masks or pixel labels (B). As an optional step, not shown in Fig. 1, the
training data set may also be augmented by rotating, flipping, blurring For pixels pi and pj , spatial colour coordinates xi and yi and xj and yj
respectively, the spatial distance is similarly defined

2
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

( ) √(̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2 ( )2̅
ds pi , pj = xi − xj + yi − yj (3)

A set of K superpixels is generated through k-means clustering,

starting with the construction of the cluster centroids.
Fig. 2 gives an example of a superpixelated image, showing a 500-
and a 50-superpixel version of an original image of platinum metal
group ore particles. As indicated by the image in the bottom right of the
Original Image 500-superpixel overlay 50-superpixel overlay figure, the ore particles in the original image (bottom, left) are still
mostly intact in the image, but texture within the particle boundaries
have been removed to processing and identification of the particles.

2.2. Labelling and preprocessing of images

In order to train neural networks for the segmentation task, a data set
with pairs of images and associated masks (labels) is required. In the
Original Image 500-superpixel image 50-superpixel image
context of deep learning, image segmentation is a pixelwise classifica­
Fig. 2. Example of a superpixelated image, showing the original image (top, tion problem, where each pixel of the original image is classified into
left), the image overlayed with 500 superpixels (top, middle) and 50 super­ two (binary) or more (multiclass) categories. Similar to cell segmenta­
pixels (top, right), as well as the original image (bottom, left), and resultant tion in the medical domain, rock segmentation is normally treated as a
500- (middle) and 50-superpixel images. binary classification problem. That is, each pixel is classified as either a
particle or non-particle. However, in this investigation, a separate class,
referred to as boundary is created, in order to force the model to focus on

Fig. 3. (a) Original image, (b) binary mask, (c) three-class mask.

Fig. 4. U-net architecture consisting of an encoder and a decoder section (after Mehrdad Yazdani – Wikipedia),

3
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

the identification of boundaries between particles.

To create the three-class data set, image annotation software called
LabelMe (https://github.com/wkentaro/labelme) was used to annotate
the image by drawing polygons along the boundaries of particles. This
produces a binary dataset, particles and non-particles, as shown in
Fig. 3a and 3b. A set of morphological operators is consequently used to
identify the boundaries of particles from the binary images, which are
then broadened by using eroding operators, as indicated in Fig. 3c).
Owing to the limited size of the data set, a range of image augmen­
tation techniques including, rotation, flipping, addition of random
Gaussian noise, and Gaussian blurring of the images were used to
expand the data set. Since the deep learning model accepted square
images, the long rectangular images had to be resized. Simple resizing
could not be used, as this could destroy the aspect ratio of the images.
Therefore, random cropping was used to crop square subimages out of
the original images. This procedure can also be seen as an effective way
to further expand the dataset. Finally, a range of morphological trans­
formations were also used to post-process the model output. More spe­
cifically, erosion followed by dilation was used to remove noise (small
holes) in the output.

2.3. U-Net architecture Fig. 5. An example of three-class ground truth mask image for Boddington ore.

The U-net architecture (Ronneberger et al., 2015), shown in Fig. 4, ( )γ

factor, γ is the focusing parameter, and 1 − ∂j is a modulating factor
essentially consists of two parts, i.e. an encoder, as shown on the left
that reduces the influence of correctly classified samples. Note that with
hand side of the diagram and a decoder on the right hand side. The
γ = 0, the focal loss is equivalent to categorical cross-entropy, and as it
encoder is typically a pretrained classification network, such as VGG16
increases, the effect of the modulating factor is likewise increased, with
or ResNet50, where an input image is encoded into a set of feature
values of γ = 2 often giving satisfactory results (Johnson and Khosh­
representations at multiple different levels. These networks are some­
goftaar, 2019).
times also referred to as dense neural networks, owing to the connec­
The dice loss component is also known as the intersection over union
tions between the encoding and decoding sections of the network. In the
loss, and is commonly used with imbalanced classes, such as associated
decoder section, the low resolution features learnt by the encoder are
with pixel classification in image segmentation problems. Likewise, the
projected onto the high resolution pixel space to get a dense classifica­
categorical cross entropy loss function is often used with multiclass
tion. To accomplish this, the decoder uses upsampling and concatena­
classification problems. Both were equally weighted in the composite
tion followed by regular convolution operations.
loss function.
With upsampling, also referred to as transposed convolution or
fractionally strided convolution, the condensed feature map is restored
3. Case study 1: Sparsely packed gold ore particles
to the original size of the input image through expansion of the feature
dimensions. This can be done by among other nearest neighbor methods,
In this case study, ore particles were sourced from a gold mine in
bilinear interpolation or transposed convolution from simplest to more
Australia. The particles had varying sizes and shapes, and were manu­
complex. With U-Net, the higher resolution feature maps from the
ally mixed to enable analysis of, for example, i) particles touching other
encoder network are also concatenated with the upsampled features to
particles in different positions, and ii) partial inclusion of small particles
enhance learning of representations with following convolutions.
in the shade of larger particles. This was a small dataset containing ten
In this investigation, a U-Net-architecture with ResNet34 as back­
training images and four independent test images only.
bone was used to build the segmentation model. In the traditional
As discussed in the previous section, the ground truth mask images
multiclass classification problem, categorical cross entropy loss is nor­
were converted from binary images to three-class images (particle,
mally used, but this loss function is not suitable for an imbalanced
boundary, and background) as shown in Fig. 5. The boundaries of par­
classification task. Compared with image pixels associated with particles
ticles (grey lines) were delineated, and a U-net model was trained to
and the non-particle background, the number of pixels associated with
segment the input image into these three classes.
particle boundaries is severely limited.
After that, the predicted masks were post-processed through
To mitigate complications that could arise from such an imbalanced
morphological transformation to fill small holes and remove noise, and
data set, the network was trained with a composite loss function, as
then a watershed algorithm was applied to obtain each separated par­
shown in eq. (4). This loss function consisted of a categorical focal loss
ticle instance as shown in Fig. 6. In this figure, each individual particle in
loss (Jcf ), as well as a dice loss function (Jd ). The focal loss function
the prediction (2nd column) is labeled with a unique colour.
compensates for the class imbalances in the pixels, given that far fewer It should be noted that the background of the images contained a
pixels belong to the particle boundaries than to particles themselves. The white lid with clear edges (left column in Fig. 6), which may cause
dice loss function is aimed at maximizing the intersection over union problems with traditional segmentation techniques. However, this was
between the ground truth mask and the prediction. not an issue for the deep learning model, which could focus on detecting
∑
m
( )γ ( ) ∑m ̂j
Lj ∩ L the edges of ore particles, while discarding other lines.
J = Jcf + Jd = − 1 − ∂j Lj log ̂
Lj − 2 (4) As suggested by a comparison of the 2nd and 3rd columns of Fig. 6,
L ∪ ̂j
L
the model could generally identify the particles reliably, even when
j=1 j=1 j

In eq. (4), L and ̂

L represent the true and predicted pixel labels for j = particles overlapped or touched.
1, 2, … m classes, respectively. The categorical focal loss is a categorical
cross entropy loss function that weights the contribution of each sample
to the loss based on the classification error, where ∂ is the weighting

4
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Fig. 6. Four Boddington HPGR test images (1st column: original input image, 2nd column: post-processed model prediction, and 3rd column: ground truth mask by
manual annotation).

4. Case study 2: PGM ores on industrial conveyor belt platinum group metal processing plant. The low resolution and the poor
illumination pose more challenges to segmenting particles than is the
This case study is based on a data set that contains videographic case under controlled laboratory conditions. Moreover, as shown in
images of ore particles obtained from an industrial conveyor belt in a Fig. 3, these ore particles are densely packed, and some boundaries

5
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Fig. 7. (left) Original input image; (right) segmentation with the watershed algorithm.

between them are barely identifiable by human visual inspection. thresholding was applied to convert images to binary images. The L2
There are 16 images in this study (10 randomly allocated to a distance measure was used to calculate the distance map of the binary
training set, and six allocated to a test set) in this study, and as shown in image.
Fig. 3, these images were obtained in an elongated rectangular shape. To Considering the mean size of particles, the local maxima (markers) in
expand the data set and preserve the aspect ratio of images, an the distance map were determined with a minimum distance of 20
augmentation technique called random cropping was used. For each pixels. Lastly, watershed was applied to segment markers from the
image, 10 different starting coordinates were randomly generated, and background.
10 sub-images were subsequently created by cropping the original An example of watershed segmentation is shown in Fig. 7. Large
image into square shape (Fig. 7, left). particles are oversegmented and generally, edge detection between
particles is poor, as these are common challenges when watershed al­
gorithms are applied to images under low illumination (Kornilov and
4.1. Watershed segmentation
Safonov, 2018).
Watershed algorithms are based on the concept of representing im­
ages as topographic surfaces with high-intensity values denoting peaks 4.2. Simple linear iterative clustering (SLIC)
and hills, and low intensity values denoting valleys. Generally speaking,
any given pixel will either be at a local minimum on this topographic In the first case study, the original image was presented to the model
surface, or would be a point where water would move to such a single that was trained to classify each pixel in the image into one of three
minimum (a catchment basin), or water would be equally likely to move classes (particle, boundary, and background). In this case study, a new
to more than one such minimum point (a watershed line). method is proposed by using simple linear iterative clustering (SLIC) to
Image segmentation is accomplished by visualizing the boundaries produce a new representation of the original image by averaging pixel
between catchment basins, if water could be allowed to rise from each values in each segment, and then training a U-net model to classify the
local minimum in the image as a topographic surface. This can give rise superpixel regions into three classes.
to oversegmentation of the image, owing to the presence of spurious The SLIC transformation process is illustrated in Fig. 8. In this study,
minima in the gradient map. Marker controlled watershed algorithms 300 segments were produced for each image, and a compactness value
control this by only allowing water to rise from designated (marked) of 10 was chosen to give more weight to spatial proximity than to colour
minima (Amankwah and Aldrich, 2010). proximity and to produce relatively square segments.
Bilateral filtering was first applied to smooth the input images by An example of the prediction of a test image is shown in Fig. 9. Most
reducing noise and sharpening edges. Morphological transformation boundaries between particles are detected reliably, and as highlighted in
was used to further reduce image noise, and after that, adaptive the red box in image on the left, the model may outperform human

Fig. 8. An example of (left) original image, (middle) SLIC boundary, and (right) representation image by averaging pixel values within each superpixel segment.

6
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Fig. 9. An example of prediction of test image using SLIC + Unet: (left) prediction,(right) ground truth.

shown in Figs. 10–12. U-net tends to slightly overestimate the equivalent

Table 1
diameter of small particles, seeing that the yellow curve is consistently
Number of segmented particles (WS: Watershed, GT: Ground Truth).
below the red ground truth curve, and on the other hand, SLIC + U-net
Method Number of Particles（#） method’s curve is more closely aligned with the ground truth curve,
Image 1 Image 2 Image 3 Image 4 Image 5 Image 6 even in the small particle size regions.
WS* 7.6 7.8 6.6 9.6 7.1 4.5
U-net 12.7 19.3 20.3 14.6 17.4 14.6 5. Conclusions
SLIC + U-net 15.1 19.6 20.8 18.6 17.5 14.5
GT** 13.8 22.6 21.7 15.4 19.9 15.4 In this paper, the application of an dense convolutional neural
*Watershed (WS), **ground truth (GT). network or U-net, was considered in the semantic segmentation of vid­
eographic images of gold ore particles in a laboratory setting, as well as
visual inspection for some challenging segmentation spots. platinum metal group ore particles on an industrial conveyor belt.
A set of quantitative measurements was used to compare the seg­ Satisfactory results could be obtained by casting the pixel classifi­
mentation performance of the watershed algorithm, three-class U-net cation problem as a three-class problem, and by preprocessing the im­
and three-class U-net + SLIC. The test set consisted of 60 images, i.e. six ages with a superpixelation method. Segregation.
original images, as well as nine augmented images for each original test of the images was significantly better than what could be obtained
image. From each group, the average number of segments was calcu­ with a basic watershed algorithm. Preprocessing of the images by use of
lated. As shown in Table 1, the watershed (WS) method mostly failed to superpixelation led to improved segmentation of smaller particles in
separate particles, as detecting particulate edges in the poorly illumi­ particular, but the differences were not large.
nated images was challenging. The study would need to be extended, as it was based on a very small
To compare particle size distributions, all test images were consid­ set of images, but the results suggest that this may be a promising
ered as a single aggregate image, of which the distribution of particle approach to the real time segmentation of densely packed particles on
areas, equivalent diameter and major axis length are plotted below as conveyor belts.

Fig. 10. Particle area distributions estimated by WS (blue line), U-net (orange line), U-net + SLIC (green line) and ground truth (red line). (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of this article.)

7
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Fig. 11. Particle equivalent diameter distribution estimated by WS (blue line), U-net (orange line), U-net + SLIC (green line) and ground truth (red line).. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 12. Particle major axis length distribution estimated by WS (blue line), U-net (orange line), U-net + SLIC (green line) and ground truth (red line). (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

CRediT authorship contribution statement Aldrich, C., Jemwa, G.T., Van Dyk, J.C., Keyser, M.J., 2010. Online analysis of coal on a
conveyor belt by use of machine vision and kernel methods. Int. J. Coal Prep. Util.
30, 331–348.
Yihao Fu: Methodology, Software. Chris Aldrich: Conceptualiza­ Al-Thyabat, S., Miles, N.J., Koh, T.S., 2007. Estimation of the size distribution of particles
tion, Supervision. moving on a conveyor belt. Miner. Eng. 20 (1), 72–83.
Amankwah, A., Aldrich, C., 2011. Estimation of particulate fines on conveyor belts by
use of wavelets and morphological image processing. Int. J. Mach. Learn. Comput. 1
(2), 132–137.
Declaration of Competing Interest Amankwah, A., Aldrich, C., 2010. Rock image segmentation using watershed with shape
markers. IEEE Applied Imagery and Pattern Recognition Workshop, Cosmos Club,
Washington DC, USA, 13-15 October 2010.
The authors declare the following financial interests/personal re­ Bai, F., Fan, M., Yang, H., Dong, L., 2021. Image segmentation method for coal particle
lationships which may be considered as potential competing interests: size distribution analysis. Particuology 56, 63–170.
Yihao Fu reports financial support was provided by Minerals Research Bamford, T., Esmaeili, K., Schoellig, A.P., 2021. A deep learning approach for rock
fragmentation analysis. Int. J. Rock Mech. Min. Sci. 145, 104839.
Institute of Western Australia.
Buades, A., Coll, B. and Morel, J.M. 2005. A non-local algorithm for image denoising. In:
Proceedings of the IEEE Computer Society Conference on Computer Vision and
Data availability Pattern Recognition (CVPR’05), San Diego, CA, USA, 20–25 June 2005; Volume 2,
pp. 60–65.
Chatterjee, S., 2013. Vision-based rock-type classification of limestone using multi-class
Data will be made available on request. support vector machine. Appl. Intell. 39, 14–27. https://doi.org/10.1007/s10489-
012-0391- 7.
Chatterjee, S., Bhattacherjee, A., Samanta, B., Pal, S.K., 2010. Image-based quality
Acknowledgements monitoring system of limestone ore grades. Comput. Ind. 61 (5), 391–408.
Chen, H., Jin, Y., Li, G., Chu, B., 2018. Automated cement fragment image segmentation
The authors wish to acknowledge financial support of this study and distribution estimation via a holistically-nested convolutional network and
morphological analysis. Powder Technol. 339, 306–313.
provided by the Minerals Institute of Western Australia via MRIWA Dan, H.-C., Bai, G.-W., Zhu, Z.-H., 2021. Application of deep learning-based image
Directors’ Scholarship M0501 (Yihao Fu). recognition technology to asphalt–aggregate mixtures: Methodology. Constr. Build.
Mater. 297, 123770.
Felzenszwalb, P.F., Huttenlocher, D.P., 2004. Efficient graph-based image segmentation.
References Int. J. Comput. Vis. 59 (2), 167–181.
Fu, Y. and Aldrich, C. 2020. Estimation of bubble size distributions in flotation froths by
Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Susstrunk, S., 2012. SLIC superpixels use of dense convolutional neural networks, IMPC 2020: XXX International Mineral
compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Processing Congress, Cape Town, South Africa, 18-22 October 2020.
Intell. 34 (11), 2274–2282. Gao, R., Sun, Z., Li, W., Pei, L., Hu, Y., Xiao, L., 2020. Automatic coal and gangue
Aldrich, C., Jemwa. G.T., Rama, M., 2009. Online analysis of coal fines on a conveyor segmentation using U-net based fully convolutional networks. Energies 13, 829.
belt by use of machine vision and kernel methods. South African Chemical https://doi.org/10.3390/en13040829.
Engineering Congress, Cape Town, Western Cape, South Africa, 20-22 September,
published on CD-ROM, ISBN: 978-1-920355-21-0.

8
Y. Fu and C. Aldrich Minerals Engineering 193 (2023) 108019

Guyot, O., Monredon, T., LaRosa, D., Broussaud, A., 2004. VisioRock, an integrated Olivier, L.E., Maritz, M.G., Craig, I.K., 2019. Deep convolutional neural network for mill
vision technology for advanced control of comminution circuits. Miner. Eng. 17 feed size characterization. IFAC-PapersOnLine 52 (14), 105–110.
(11–12), 1227–1235. Olivier, L.E., Maritz, M.G., Craig, I.K., 2020. Estimating ore particle size distribution
Hamzeloo, E., Massinaei, M., Mehrshad, N., 2014. Estimation of particle size distribution using a deep convolutional neural network. IFAC-PapersOnLine 53 (2),
on an industrial conveyor belt using image analysis and neural networks. Powder 12038–12043.
Technol. 261, 185–190. Patel, A.K., Chatterjee, S., Gorai, A.K., 2017. Development of machine vision-based ore
Jemwa, G.T., Aldrich, C., 2012. Estimating size fractions of coal particles on conveyor classification model using support vector machine (SVM) algorithm. Arabian J.
belts using image texture modelling methods. Expert Syst. Appl. 39, 7947–7960. Geosci. 10, 1–16. https://doi.org/10.1007/s12517-017-2909-0.
Jemwa, G.T. and Aldrich, C. 2010. Estimating particle size fractions from image data: A Qin, W., Han, F., Yuan, Y., Zhao, W., Ibragimov, B., Gu, J., Xing, L., 2018. Superpixel-
multiscale approach using multiple kernel learning. In: Proceedings of the XXV based and boundary-sensitive convolutional neural network for automated liver
International Mineral Processing Congress, Brisbane, Australia, 6-10 September segmentation. Phys. Med. Biol. 63 (9), 095017.
2010. Qiu, Z., Dou, D., Zhou, D., Yang, J., 2021. On-line prediction of clean coal ash content
Johnson, J.M., Khoshgoftaar, T.M., 2019. Survey on deep learning with class imbalance. based on image analysis. Measurement 173, 108663.
J. Big Data 6, 27. https://doi.org/10.1186/s40537-019-0192-5. Ren, X. and Malik, J. 2003. Learning a classification model for segmentation. In:
Kaartinen, J., Tolonen, A., 2008. Utilizing 3D height measurement in particle size International Conference on Computer Vision, 10–17.
analysis. IFAC Proceedings Volumes 41 (2), 3292–3297. Ronneberger, O., Fischer, P. and Brox, T. 2015. U-Net: Convolutional networks for
Karimpouli, S., Tahmasebi, P., 2019. Segmentation of digital rock images using deep biomedical image segmentation. arXiv:1505.04597.
convolutional autoencoder networks. Comput. Geosci. 126, 142–150. Schenk, F., Tscharf, A., Mayer, G. and Fraundorfer, F. 2019. Automatic muck pile
Kistner, M., Jemwa, G.T., Aldrich, C., 2013. Monitoring of mineral processing systems by characterization from UAV images. SPRS Annals of the Photogrammetry, Remote
using textural image analysis. Miner. Eng. 52, 169–177. Sensing and Spatial Information Sciences, IV-2/W5: 163-170.
Kornilov, A.S., Safonov, I.V., 2018. An overview of watershed algorithm Shi, J., Malik, J., 2000. Normalized cuts and image segmentation. IEEE Trans. Pattern
implementations in open source libraries. Journal of Imaging 4, 123. Anal. Mach. Intell. 22 (8), 888–905.
Kulatilake, P.H.S.W., Qiong, W., Hudaverdi, T., Kuzu, C., 2010. Mean particle size Singh, V., Rao, S.M., 2005. Application of image processing and radial basis neural
prediction in rock blast fragmentation using neural networks. Eng. Geol. 114 (3–4), network techniques for ore sorting and ore classification. Miner. Eng. 18,
298–311. 1412–1420.
Leiva, C., Acuña, C., Castillo, D., 2021. Development and validation of an online analyzer Sousa, R., Futuro, A., Fiúza, A., Leite, M.M., 2020. Pre-concentration at crushing sizes for
for particle size distribution in conveyor belts. Minerals 11 (6), 581. https://doi.org/ low-grade ores processing – ore macro texture characterization and liberation
10.3390/min11060581. assessment. Miner. Eng. 147, 106156.
Leroy, S., Dislaire, G., Bastin, D., Pirard, E., 2011. Optical analysis of particle size and Stutz, D., Hermans, A. and Leibe, B. 2017. Superpixels: An evaluation of the state-of-the-
chromite liberation from pulp samples of a UG2 ore regrinding circuit. Miner. Eng. art. arXiv:1612.01601v3 [cs.CV] 19 Apr 2017.
24 (12), 1340–1347. Thurley, M.J., 2009. Automated online measurement of particle size distribution using
Li, H., Asbjörnsson, G. and Lindqvist, M. 2021. Image process of rock size distribution 3D range data. IFAC Proceedings Volumes 42 (23), 134–139.
using DexiNed-based neural network. Minerals, 11, 736. https:// doi.org/10.3390/ Thurley, M.J., 2011. Automated online measurement of limestone particle size
min11070736. distributions using 3D range data. J. Process Control 21 (2), 254–262.
Liu, Y., Jiang, P.-T., Petrosyan, V., Li, S.-J., Bian, J., Zhang, L. and Cheng, M-M. 2018. Tomasi, C. and Manduchi, R. 1998. Bilateral filtering for gray and color images. In
DEL: Deep embedding learning for efficient image segmentation. In: Proceedings of Proceedings of the Sixth International Conference on Computer Vision, Bombay,
the Twenty-Seventh International Joint Conference on Artificial Intelligence, 864- India, 7 January, 839– 846.
870, Stockholm, Sweden, 13-19 July https://doi.org/10.24963/ijcai.2018/120. Varfolomeev, I., Yakimchuk, I., Safonov, I., 2019. An application of deep neural networks
Liu, X., Zhang, Y., Jing, H., Wang, L., Zhao, S., 2020. Ore image segmentation method for segmentation of microtomographic images of rock samples. Computers 8 (4), 72.
using U-Net and Res_Unet convolutional networks. RSC Adv. 10 (16), 9396–9406. Wang, W., Li, Q., Xiao, C., Zhang, D., Miao, L., Wang, L., 2021. An improved boundary-
Liu, Y., Zhang, Z., Liu, X., Wang, L., Xia, X., 2021a. Ore image classification based on aware U-net for ore image semantic segmentation. Sensors 21, 2615.
small deep learning model: Evaluation and optimization of model depth, model Wu, X., Liu, X., Duan, J., 2021. Online size distribution measurement of dense iron green
structure and data size. Miner. Eng. 107020. pellets using an efficient and multiscale nested U-net method. Powder Technol. 387,
Liu, Y., Zhang, Z., Liu, X., Wang, L., Xia, X., 2021b. Deep learning-based image 584–600.
classification for online multi-coal and multi-class sorting. Comput. Geosci. 157, Yang, F., Sun, Q., Jin, H. and Zhou, Z. 2020. Superpixel segmentation with fully
104922. convolutional networks. In: 2020 IEEE/CVF Conference on Computer Vision and
Long, J., Shelhamer, E. and Darrell, T. 2014. Fully convolutional networks for image Pattern Recognition (CVPR), pp. 13961-13970, doi: 10.1109/
segmentation. arXiv:1411.4038. CVPR42600.2020.01398.
Manana, R., Artieda, J.J., Catalina, J.C., 1985. Ore sorting and artificial vision. IFAC Zhang, Z., Yang, J., Ding, L., Zhao, Y., 2012. Estimation of coal particle size distribution
Proceedings Volumes 18 (6), 235–240. by image segmentation. Int. J. Min. Sci. Technol. 22 (5), 739–744.
Min, H., Peipei, W., Huihui, J., 2012. Recognition of coal and stone based on SVM and Zhang, Z., Yang, J., Su, X., Ding, L., Wang, Y., 2013. Multi-scale image segmentation of
texture. J. Comput. Eng. Des. 33, 1117–1121. https://doi.org/10.3969/j.issn.1000- coal piles on a bely based on the Hessian matrix. Particuology 11 (5), 549–555.
7024.2012.03.055. Zhang, Z., Liu, Y., Hu, Q., Zhang, Z., Wang, L., Liu, X., Xia, X., 2020. Multiinformation
Mu, X., Qi, H., Li, X., 2020. Automatic segmentation of images with superpixel similarity online detection of coal quality based on machine vision. Powder Technol. 374,
combined with deep learning. Circuits Systems Signal Process. 39 (2), 884–899. 250–262.
Mukherjee, D.P., Potapovich, Y., Levner, I., Zhang, H., 2009. Ore image segmentation by
learning image and shape features. Pattern Recogn. Lett. 30 (6), 615–622.