2021 International Conference on Nascent Technologies in Engineering (ICNTE 2021)

Machine Learning based Image Processing for Iron

Ore Pellet Size Analysis

Arya Jyoti Deo1,2, Animesh Sahoo3, Santosh Kumar Behera1,2, Debi Prasad Das1,2
1
CSIR-Institute of Minerals and Materials Technology, Bhubaneswar, India
2
Academy of Scientific and Innovative Research (AcSIR), Ghaziabad, India
3
Birla Institute of Technology and Science Pilani, Goa Campus, India
2021 4th Biennial International Conference on Nascent Technologies in Engineering (ICNTE) | 978-1-7281-9061-7/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICNTE51185.2021.9487768

Email: [email protected], [email protected], [email protected], [email protected]

Abstract—Image processing based system have been proposed do not perform identically, size analysis of pellets at individual
for size analysis of iron ore pellets in real-time. One such disc is very much essential for timely controlling the disc
commercial system named as Eye-On-Pellet is introduced by parameters such as disc speed, feed rate and water addition to
CSIR-IMMT, India which is being used in pelletization industries. operate the disc in optimum sized pellet production condition.
Further improvement in this system by using machine learning
algorithms is proposed in this paper. Two locations of image Conventionally, multiple human operators are engaged
capture to derive pellet size information is dealt in this work. Eye- to visually observe each disc at different time interval and find
On-Pellet system captures the pellet size information from the disc out the disturbance in the pellet production and then control the
discharge point which uses traditional image processing disc parameters. Since the pellets are formed from dusty iron ore
algorithms. In addition to this, a new approach is proposed where fines, the pelletization plant area remains very much dusty.
the image of pellets inside the disc can be used to qualitatively Human operator working in such harsh dusty environment may
know if the size is small, big or normal using convolutional neural suffer from health issues [9].
network.
Therefore, camera and image processing based system
Keywords— Iron ore pellet; size analysis; image processing; to watch the pelletization discs and generate size analysis report
machine learning; convolutional neural network; Eye-On-Pellet and trend has been developed by CSIR-Institute of Minerals and
Materials Technology, Bhubaneswar, India. The name of the
I. INTRODUCTION system is “Eye-On-Pellet” [10]. This has been adopted by some
Use of steel is increasing day by day in developing countries pelletization plants in India and the performance is satisfactory.
due to new infrastructure development [1]. Steel is produced Computer vision and advanced instrumentation have
mostly in blast furnaces and direct reduced iron (DRI) plants. been applied to many industrial problems in past. On-line iron-
Blast furnace requires lumpy and high-grade iron ore. However, ore slurry monitoring for real-time process control of pellet
low-grade iron ores undergo beneficiation process to improve making processes using laser-induced breakdown spectroscopy
the grade. During beneficiation process the iron ores are ground has been proposed [11]. An industrial 3D vision system for size
to fine size, called as iron ore fines. Mining of iron ores also measurement of iron ore green pellets using morphological
produces a large number of fines. These fines cannot be directly image segmentation has also been proposed [12]. These are
used in blast furnace for steel making. To utilize the iron ore expensive solutions. Discrete element method based modeling
fines in blast furnace and direct reduced iron plants, fines are and simulation study has been reported for green iron ore pellet
converted to a sized spherical balls called pellets and the process classification using roller screen [13]. In-line image-based
is called pelletization [2-7]. Typical required size range of particle size measurement of hot-melt extruded pellets has been
pellets are 8 to 18mm. Using mechanical screening [7], the under reported [14]. Multivariate video analysis and Gaussian process
size and over size pellets are separated from the total produced regression model based soft sensor has been proposed for online
pellets. Higher amount of under and over sized pellet production estimation and prediction of nickel pellet size distributions [15].
reduces the total production leading to loss of production and In another publication, artificial vision system for particle size
increases the recycling load [8]. In fact, during pelletization in a characterization from bulk materials has been reported [16].
disc pelletizer, the size of the pellets does not remain uniform Image processing based size characterization of green pellets on
and change due to variations in the characteristics of feed iron a conveyor has been reported [17]. Detection and segmentation
ore fines. Pelletizer discs may sometimes produce higher of iron ore green pellets in images captured at disc discharge
undersize pellets and some other times produce higher oversized using lightweight U-net deep learning network has been recently
pellets. This is mainly due to the feed characteristics such as feed proposed [18]. This shows that not only use of image processing
rate variation, moisture content and particle size distribution of algorithms but also use of advanced deep learning algorithms
feed fines. Since there are at least six to ten such pelletizer discs have been used for iron ore pellet size analysis.
and only one screen to separate the under and over sized pellets,
by using the mechanical screening method size analysis of While working on such a problem of accurately and
individual discs cannot be done. As all pelletizer discs in a plant timely size analysis of iron ore green pellets during its

978-1-7281-9061-7/21/$31.00 ©2021 IEEE

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on May 17,2023 at 14:34:16 UTC from IEEE Xplore. Restrictions apply.
 2021 International Conference on Nascent Technologies in Engineering (ICNTE 2021)

II. EYE-ON-PELLET SYSTEM

Eye-On-Pellet is a patented technology for online pellet size
monitoring system developed by CSIR-IMMT Bhubaneswar
which works on the principle of image processing. A camera is
installed near the disc and is focused to the disc discharge point
where the images of falling pellets are being captured and
analyzed in real-time to get size distribution of pellets in a
continuous manner. In this case, the pellets are generally imaged
in a monolayer basis and the size analysis can be done using
traditional image processing techniques. In this case, the
background plays an important role due to which detection of
pellets and its accurate size analysis is a challenge. Since the
background may change with time due to dust accumulation and
(a)
maintenance failure, the performance of the size analysis may be
degraded. Some representative examples of size analysis using
image processing result is shown in Fig. 1 (a) (raw Image) and
1 (b) (processed image). A representative size analysis trend
obtained from a pelletization process by use of the Eye-On-
Pellet is shown in Fig. 1 (c) where the quantitative size analysis
is possible and increasing and decreasing trend is clearly
identifiable.
III. DEEP LEARNING BASED SIZE ANALYSIS
In this work, it is assumed that a camera is fixed to view the
inside of a disc pelletizer, where a portion of the disc pelletizer
is imaged. In such an image, there is no distinction between
background and object. Fine iron ores mixed with binder are fed
(b) by a weigh feeder at a feed rate of 80 to 140 TPH. Water is

(c)

Fig. 1. (a) Original image captured by the Eye-On-Pellet Camera (b)

processed image with size measurement of individual pellets by Eye-On-Pellet
algorithm (c) size analysis result

production at disc pelletizer, a machine learning algorithm is

used in this paper to provide an alternate size analysis technique.
Machine learning algorithms have been used recently for many
industrial applications. Deep learning (DL) using convolutional
neural network (CNN) has been an important breakthrough in
the area of image processing. In this work, the same has been
also attempted for development of pellet size analysis system. It
has been shown that by use of CNN, the pellet size classification Fig. 2. The Weighted Average Ensembled Architecture used for the
can be done without individually counting and sizing the pellets. classification task.

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on May 17,2023 at 14:34:16 UTC from IEEE Xplore. Restrictions apply.
 2021 International Conference on Nascent Technologies in Engineering (ICNTE 2021)

sprayed and the disc is rotated at about 7 to 10 RPM, which can resolution 494 659. The images in the dataset are divided into
be varied to control the size of the pellets. The disc is kept at an four classes, i.e. Big Size, Medium Size, Small Size and Powder
angle so that the fines are moved up due to rotational motion of Material, each containing 500 images. For each class, 400
the disc and after reaching certain height these fines fall. During images are used for training purpose and 100 images are used
this motion they combine with other particles and the size grows. for testing. Figure 3 shows example images of each class. Data
If one expert operator observes the inside portion of the disc, he augmentation is used to generate more number of training
can tell, if the size is bigger or smaller and also he can tell if the samples as required for transfer learning of the pre trained
pellet size is growing or reducing. networks.
This concept is used to build an artificial intelligence based C. Training Setup
system to find the pellet growth condition inside the disc. All the experiments and analysis are carried out in Google
Specifically an ensemble deep learning network framework is Colaboratory platform. Python 3.6.9 is used as the programming
newly proposed in this paper for pellet size classification task. environment. Keras 2.4.3 is used to import the baseline models
A. Proposed Deep Learnning Algorithm and to perform transfer learning operation. Wrapping and
rescaling of the images are done for creating the training and
The network architecture used in this paper ensembles three
testing generators. The size of images is rescaled to 224 224
popular CNN models, VGG16[19], Mobilenet [20] and
to match with the input size of the pre trained models. 25% of
Resnet50[21], pre trained on ImageNet [22] dataset. Weighted
the augmented training images are used for validation purpose.
ensembling method used in this paper, provides higher weight
All individual models and the proposed weighted average
to the model with better performance. Also averaging helps in
ensembled(WAE) model are trained with Adam optimizer for 20
reducing the probability of overfitting of training data. For the
epochs at a batch size of 80 with an initial learning rate of 10-3.
sake of applying transfer learning to each of these models, all
weights parameters except those of last few layers are freezed. D. Results and Discussion
Specifically, the fully connected layers of each models are Figure 4 shows the accuracy curve of all the models during
allowed to learn new weights in order to make our classification training and validation. It can be observed that the accuracy of
task reliable. Additionally, the 12th and 13th convolutional layer VGG16 [19] is about 70% and that of Mobilenet [20] is around
of Mobilenet [20] are also unfreezed to get the weights updated 90%. The performance of Resenet50 [21] and the proposed
during training operation. A dropout layer is used at the output WAE architecture are in acceptable range. The validation
of every individual network to avoid overfitting. The output of accuracy of the WAE model outperforms the others. The values
all these models are then ensembled using weighted averaging of individual confusion matrix are tabulated for all models in
method [23]. Batch normalization layers are used before the Table-I for comparison purpose. It can be seen that the all
final fully connected layer to ensure efficient performance. The models are able to classify the powder materials successfully.
detailed network architecture is shown in Fig. 2. Considering the individual performance, VGG16 [19], failed to
B. Data Generation classify the medium sized pellets from the small size category.
However, the WAE architecture outperforms all the individual
As there is no public dataset available for the mentioned
performance. It can be further verified from the performance
purpose, a dataset containing 2000 images of pellets and powder
comparison in Table-II, that the WAE model is better than
materials is created for this purpose. The captured images are of

Fig. 4. Comparison of training and validation accuracy of individual models

Fig.3. Example images from the prepared dataset representing four classes, i.e.
and weighted average ensembled model.
Pellets of Big Size, Medium Size, Small Size and Powder Materials.

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on May 17,2023 at 14:34:16 UTC from IEEE Xplore. Restrictions apply.
 2021 International Conference on Nascent Technologies in Engineering (ICNTE 2021)

TABLE I. COMPARATIVE CONFUSION MATRIX VALUES OF INDIVIDUAL AND WAE MODELS

Pellet Big Size Medium Size Small Size Powder

Categories VGG Mob- Res- WAE VGG Mob- Res- WAE VGG Mob- Res- WAE VGG Mob- Res- WAE
16 ilenet net50 model 16 ilenet net50 model 16 ilenet net50 model 16 ilenet net50 model
Big Size 0.98 0.98 1 1 0.01 0.02 0 0 0 0 0 0 0.01 0 0 0
Medium
0.01 0 0 0 0.6 0.96 0.99 1 0.39 0.04 0.01 0 0 0 0 0
Size
Small
0 0 0 0 0.06 0.17 0.02 0.01 0.94 0.83 0.98 0.99 0 0 0 0
Size
Powder 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1

TABLE II. PERFORMANCE INDICATOR COMPARISON FOR THE INDIVIDUAL AND WAE MODELS

VGG16 Mobilenet Resnet50 WAE model

Pellet
Categories Preci F1- Preci F1- Preci F1- Preci F1-
Recall MCC Recall MCC Recall MCC Recall MCC
sion Score sion Score sion Score sion Score

Big
0.99 0.98 0.99 0.98 1 0.98 0.99 0.99 1 1 1 1 1 1 1 1
Size
Medium
0.91 0.6 0.72 0.67 0.85 0.96 0.9 0.86 0.98 0.99 0.99 0.98 0.99 1 1 0.99
Size
Small
0.69 0.94 0.79 0.74 0.95 0.83 0.88 0.86 0.99 0.98 0.99 0.98 1 0.99 1 0.99
Size
Powder 0.99 1 1 0.99 1 1 1 1 1 1 1 1 1 1 1 1

individual network performance. Precision, Recall and F1-Score REFERENCES

and Matthews correlation coefficient (MCC) are used as the
[1] National Steel Policy(NSP), 2017, https://steel.gov.in/national-steel-
evaluation benchmarks in this case. policy-nsp-2017 (last seen 7th Feb 2020).
[2] P. J. Kakela, "Iron ore: Energy, labor, and capital changes with
IV. CONCLUSION technology." Science 202, no. 4373 pp. 1151-1157, 1978.
In this paper, an industrial application of image processing [3] D. Zhu, J. Pan, L. Lu, and R. J. Holmes. "Iron ore pelletization." Iron Ore,
with and without machine learning is presented. One Woodhead Publishing, pp. 435-473. 2015.
commercial system developed by CSIR-IMMT, India which [4] G. P. Singh, R. P. Choudhary, H. Vardhan, M. Aruna, and A. B. Akolkar.
solely depends on traditional image processing is presented "Iron ore pelletization technology and its environmental impact
assessment in eastern region of India-a case study." Procedia Earth
where quantitative analysis is quite impressive. Results of an Planetary Sci 11 pp. 582-597, 2015.
actual plant data of size analysis showing increasing and
[5] L. Wenquan, "Knowledge and view on development of pellet production
decreasing trend has been shown. Another simple method using in China [J]." Ironmaking 3 4, 2006.
the concept of visual perception of a human supervisor to derive [6] S. L. De Moraes, and T. R. Ribeiro. "Brazilian iron ore and production of
the information about big or small sized pellet is also proposed. pellets." Mineral Processing and Extractive Metallurgy Review 40, no. 1,
Weighted average ensembling of individual CNN models is pp. 16-23, 2019.
applied to classify the images based on the pellet size in the [7] S. K. Kawatra, "New Paradigms for Iron Ore Pelletization," Extraction
image. During testing for a simulated condition, the values of 2018, pp. 953-959. Springer, Cham, 2018.
the evaluation metrics suggest that performance of Weighted [8] M. Athayde, S. F. Nunes, and M. C. Bagatini "A case study of pellet size
average ensembling method is better than prediction obtained fraction influence on pelletizing operation." Mineral Processing and
Extractive Metallurgy Review 39, no. 4 pp. 276-283, 2018.
through individual models. Under such conditions, it needs
[9] L. P. C. Faria, and C. J. C. Blanco. "Dust emission in iron ore storage
about 30 to 40 minutes to complete a whole training procedure. yard: an environmental and occupational approach." Research, Society
After training, the weight model is obtained, and it needs few and Development 9, no. 2, p. 05921873, 2020.
seconds to process one test patch image. The computing time of [10] Report on “Eye-On-Pellet”, CSIR-IMMT, Bhubaneswar, India
the proposed method can be further reduced by employing http://www.immt.res.in/TechProfile/Proforma%20for%20EYE%20%20
computer with higher configurations and is prosperous for real- ON%20PELLET%20Technologies.pdf (last seen 7th Feb 2020).
time measurement. [11] L. Barrette, and S. Turmel. "On-line iron-ore slurry monitoring for real-
time process control of pellet making processes using laser-induced
ACKNOWLEDGMENT breakdown spectroscopy: graphitic vs. total carbon detection,"
Spectrochimica Acta Part B: Atomic Spectroscopy 56, no. 6, pp. 715-723,
S. K. Behera and D. P. Das acknowledge Instrumentation 2001.
Development Program of DST, Govt of India for financial [12] M. J. Thurley, and T. Andersson. "An industrial 3D vision system for size
support, A.J. Deo acknowledges CSIR, India for fellowship measurement of iron ore green pellets using morphological image
under CSIR- SRF scheme. segmentation," Minerals engineering 21, no. 5 pp. 405-415, 2008.

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on May 17,2023 at 14:34:16 UTC from IEEE Xplore. Restrictions apply.
 2021 International Conference on Nascent Technologies in Engineering (ICNTE 2021)
[13] E Silva, B. Barbabela, E. R. da Cunha, R. M. de Carvalho, L M. Tavares. [18] J. Duan, X. Liu, X. Wu, C. Mao. "Detection and segmentation of iron ore
“Modeling and simulation of green iron ore pellet classification in a single green pellets in images using lightweight U-net deep learning network."
deck roller screen using the discrete element method,” Powder technology Neural Computing and Applications, pp. 1-16, 2019.
332, pp. 359-370, 2018. [19] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
[14] D. Treffer, P. R. Wahl, T. R. Hörmann, D. Markl, S. Schrank, I. Jones, P. large-scale image recognition,” 3rd International Conference on Learning
Cruise et al. "In-line implementation of an image-based particle size Representations, ICLR 2015.
measurement tool to monitor hot-melt extruded pellets," International [20] A. G. Howard et al., “MobileNet V1,” arXiv Prepr. arXiv1704.04861,
journal of pharmaceutics 466, no. 1-2 pp. 181-189, 2014. 2017
[15] J. J. Y. Chen, and Y. Zhang. "Multivariate video analysis and Gaussian [21] K. He, X. Zhang, S. Ren, J. Sun. "Deep residual learning for image
process regression model based soft sensor for online estimation and recognition." In Proceedings of the IEEE conference on computer vision
prediction of nickel pellet size distributions," Computers & Chemical and pattern recognition, pp. 770-778. 2016.
Engineering, 64 pp. 13-23, 2014.
[22] A. Krizhevsky, I. Sutskever, and H. Geoffrey E., “Imagenet,” Adv. Neural
[16] P. Facco, A.C. Santomaso, M. Barolo, Artificial vision system for particle Inf. Process. Syst. 25, 2012, doi: 10.1109/5.726791.
size characterization from bulk materials, Chemical Engineering Science
(2017),doi:http://dx.doi.org/10.1016/j.ces. 2017.01.053 [23] D. Jimenez, "Dynamically weighted ensemble neural networks for
classification," 1998 IEEE International Joint Conference on Neural
[17] V. Subramanyam, P. Patra, and M. K. Singh. “Automatic image Networks Proceedings. IEEE World Congress on Computational
processing based size characterization of green pellets,” International Intelligence (Cat. No.98CH36227), Anchorage, AK, 1998, pp. 753-756
Journal of Automation and Smart Technology 7, no. 3 pp. 85-91, 2017. vol.1, doi: 10.1109/IJCNN.1998.682375.

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on May 17,2023 at 14:34:16 UTC from IEEE Xplore. Restrictions apply.