Professional Documents
Culture Documents
PII:
DOI:
Reference:
S0098-3004(15)00091-6
http://dx.doi.org/10.1016/j.cageo.2015.04.008
CAGEO3532
Petroleum Exploration Department, Faculty of Mining Engineering, University College of Engineering, University of Tehran,
McGill Center for Bioinformatics, School of Computer Sciences, McGill University, Trottier Building, Room#: 3140, 3630
Department of Computer Engineering, Faculty of Electrical and Computer Engineering, University of Birjand, Birjand, Iran.
Department of Mining Engineering, Faculty of Engineering, University of Birjand, Birjand, Iran. namehran@birjand.ac.ir.
Abstract
Minerals segmentation in thin sections is a challenging, popular, and important research topic in
computational geology, mineralogy, and mining engineering. Mineral segmentation in thin sections
containing altered minerals, in which there are no evident and close boundaries, is a rather complex
process. Most of the thin sections created in industries include altered minerals. However, intelligent
segmentation in thin sections containing altered minerals has not been widely investigated in the literature,
and the current state of the art algorithms are not able to accurately segment minerals in such thin sections.
In this paper, a novel method based on incremental learning for clustering pixels is proposed in order to
segment index minerals in both thin sections with and without altered minerals. Our algorithm uses twelve
color features that are extracted from thin section images. These features include: red, green, blue, hue,
saturation and intensity, under plane and cross polarized lights in maximum intensity situation. The
proposed method has been tested on 165 igneous samples and the overall accuracy of 92.15% and 85.24%
*
has been obtained for thin sections without altered minerals and thin sections containing altered minerals,
respectively. Experimental results indicate that the proposed method outperforms the results of other
similar methods in the literature, especially for segmenting thin sections containing altered minerals. The
proposed algorithm could be applied in applications which require real time segmentation or efficient
identification map such as petroleum geology, petrography and NASA's Mars explorations.
Keywords: Intelligent mineral segmentation, Thin sections, Incremental learning, RGB and HSI color
features, Image processing.
1. Introduction
The main purpose of studying thin section images is the segmentation of minerals (Fabbri, 1984; Allard
and Benn, 1989; Simigian and Starkey, 1989). Intelligent mineral segmentation in thin sections is a crucial
step for performing further processing on minerals such as minerals identification and naming rocks (Izadi
et al., 2013). In addition, it makes possible some quantitative measurements such as grain size and shape
analysis, which could be used to address a wide variety of geological problems (Goodchild and Fueten,
1998). Therefore, developing an algorithm for minerals segmentation in thin sections is an important step
in computational geology.
In recent years, some scholars have developed various algorithms and methods using color features for
minerals segmentation in thin sections (Heilbronner, 2000; Van den berg et al., 2002). Most of these
methods are based on color edge detection; however, this technique is not a perfect method and only works
well on clean materials like ceramics (Arnould et al., 2001). As one of the first attempts for segmenting
minerals in thin sections, Goodchild and Fueten (1998) proposed an algorithm based on such image
processing techniques as Gaussian smoothing, gradient calculation, thresholding and image
skeletonisation. Boundary detection of grains in thin sections based on gradient filtering and using Lazy
Grain Boundary (LGB) method was proposed by Heilbronner (2000). LGB method, which is public
domain image processing software, consists of a set of macro commands programmed for NIH images. In
2
another study, Van den Berg et al. (2002) suggested an alternative algorithm for boundary detection of
grains in digital images of thin sections by using maximum likelihood classifier. Another algorithm for
grains boundary detection in thin sections was developed by Zhou et al. (2004-A). They used image
processing techniques based on phi-image (rotation stage of petrographic microscope) and max-image
(maximum intensity of minerals) as the - methodology of their proposed algorithm. In another study,
(Zhou et al., 2004-B) developed a method based on edge finding and region growing techniques for
minerals segmentation. In their study, the color edges were created by combining results of edge operators
and a color edge detection algorithm, and then a region growing algorithm was used for image
segmentation. Another method based on the watershed algorithm for minerals segmentation in thin
sections was proposed by Barraud (2006). After applying the watershed algorithm, the map of the grain
boundaries was corrected manually. In another study, Obara (2007) proposed an algorithm based on color
system transformation, and image processing techniques. The watershed algorithm was also used for grain
segmentation in RGB, CIELab, CIELuv, YIQ and YUV color spaces. Fueten and Mason (2007) developed
an algorithm based on the method proposed by Goodchild and Fueten (1998). Moreover, they used
artificial neural network (ANN), and color and texture features for mineral segmentation. They used ANN
to also examine the edges of minerals. Their approach could significantly reduce the amount of manual
corrections after automatic edge detection process. Boundary detection of minerals using the level set
method (LSM) was proposed by Lu et al. (2009). LSM consists of two steps which are initializing the
curve and then evolving it by solving the level set equation. In another study, Yesiloglu-Gultekin et al.
(2012) suggested a computer program (TSecSoft) which segmented minerals by using minimum
birefringence color intensity and several pre-processing techniques. It creates a final segmentation map by
user feedback, and then determines the mineral percentages. Furthermore, Filho et al., (2013) developed a
semiautomatic algorithm based on image foresting transform (IFT) for minerals segmentation in thin
sections.
Most minerals in thin sections existing in industries such as petroleum geology and petrology are altered
minerals, in which mineral boundaries are not clearly evident and close. Segmentation of thin sections
containing altered minerals is a very difficult problem, and it also seems nearly impossible through
methods formerly used in the literature (Arnould et al., 2001). So far, a general method for minerals
segmentation in thin section containing altered minerals has not been proposed, and the current state of the
art techniques are unable to obtain high accuracy for minerals segmentation in such thin sections.
Therefore, developing an online and real time algorithm for minerals segmentation in thin sections
containing altered minerals could be of great help in carrying out research and in mining industries. In this
paper, using color features, a new intelligent method for minerals segmentation in thin sections is proposed
based on an incremental clustering algorithm. While clustering based segmentation is a new approach to
mineral segmentation, most of clustering algorithms are unable to cluster all the minerals in thin sections.
This is because they need the number of clusters (minerals) as the input parameter. However, it is
impossible to anticipate the number of minerals in thin sections. Incremental clustering algorithm is a
method of clustering that does not need to determine the number of clusters (minerals) as the input
parameter, and it is able to determine the number of clusters in incoming data incrementally (Sadri et al.,
2006). The proposed method is able to efficiently segment minerals in thin sections both with and without
altered minerals with high accuracy.
The rest of this paper is organized as follow: in Section 2, mineralogy background information is
described. In Section 3, the database collection and details of images are presented. In Section 4, the
proposed intelligent method for minerals segmentation in thin section is introduced. Then, in Section 5,
experimental results of the proposed method are discussed, and finally, in Section 6, conclusions are drawn
and suggestions for further works are presented.
In order to examine the mineralogical properties of minerals in thin sections, a polarizer rotation stage or a
polarizer microscope (Fig. 1) is used, which works based on polarized light (Dyar et al., 2008). Optical
mineralogy is a branch of mineralogy in which thin sections are studied by polarizer microscopes with
rotation stage.
Fig. 1. A polarizer microscope with rotation stage. The light source (B) is embedded on a stand (A) and the light ray passes
through the polarizer (C). Thin sections are put on the rotation stage for examination (D) and their components arewatched by
objectives (E) and oculars (G). Analyzer (F) converts plane polarized light into cross polarized light.
In order to study rocks, some slices are turned into thin sections by special mineralogy techniques. Thin
sections are a 30m thick slice of rocks mounted on a glass microscope slide (Zhou et al., 2004). Minerals
in thin sections have specific optical properties which are different in plane and cross polarized lights.
Most minerals in the plane polarized light are colorless (Fig. 3). Hence, cross polarized light is used too for
studying minerals in thin sections.
3. Database Collection
To create a database for this study, images were captured in a maximum intensity situation by a digital
camera using a polarizer microscope with rotation stage at plane polarized and cross polarized lights. The
maximum intensity situation is defined as the maximum interference color that every pixel attains during a
180 degree rotation of the polarizers (Ross et al, 2001). In this study, 165 igneous thin sections consisting
of 28 igneous common minerals were collected (Table 1), and they were used to develop the proposed
algorithm. From each thin section, two images were captured in RGB color space under plane and cross
polarized lights with the size of 300250 pixels and the resolution of 96 dpi.
Table 1. Number of minerals samples used in this study
Number of
Row#
Number of
Mineral
Row#
Mineral
samples
samples
Amphibole
16
15
Leucite
Analcime
16
Chlorite
Andalusite
17
Muscovite
Actinolite
18
Nepheline
Apatite
19
Nosean
Biotite
20
Olivine
Cancrinite
21
Orthoclase
10
Chabazite
22
Plagioclase
Chloritoid
23
Pyroxene
12
10
Clinozoisite
24
Quartz
11
11
Enstatite
25
Sanidine
12
Kyanite
26
Sphene
13
Epidote
27
Topaz
14
Garnet
28
Zircon
subsections.
6
Pre-processing
steps
Extracting
color features
Clustering
clustering algorithm is applied for intelligent mineral segmentation. The output of the proposed method includes several black
and white images in which the index minerals are shown.
4.1 Pre-processing
In mineralogy laboratory, images are captured by different microscopes. Therefore, images are obtained
with various sizes and resolutions. The size of images is an important parameter because the running time
of the algorithm depends on it. Therefore, a general size should be determined. In this paper, after several
evaluations through trial and error, images are resized into 200150 pixels. In addition, a median filter
with the 33 mask size is used for smoothing the images and reducing their noise (Fig. 3).
Fig. 3. Applying median filter mask size 33. A: original image and B: filtered image
In order to give the same values to the features, the RGB and HSI components are linearly mapped
between 0 and 1 according to (Eq. 1) (Rashedi et al., 2009). Furthermore, histogram equalization is used
for enhancing the contrast of images (Fig. 4)
Vinput Vmin
eps
Vmax Vmin
(1)
CR
CG
CB
DR
DG
DB
Fig. 4. Histogram equalization. A: low contrast image, B: improved image contrast using histogram equalization, histogram of
low contrast image for Red (CR), green (CG) and blue (CB) components, histogram of improved contrast image for Red (DR),
green (DG) and blue (DB) components
converted into the HSI color space (for equations converting RGB color space to HSI color space, refer to
Gonzalez et al. (2009). Afterwards, the feature vector which is assigned to each pixel in thin section
images is a concatenation of 12 features including both RGB and HSI features in plane and cross polarized
lights. A schematic diagram of the feature vector is shown in Fig. 5. The details of these features are
presented in the next two subsections.
HSI color space
Fig. 5. The feature vector which is extracted from each pixel in the thin section images. It consists of six features from plane and
six features from cross polarized images
4.2.1
RGB color space is an MN3 array of color pixels, where each color pixel is a triplet containing red,
green and blue components of an RGB image at a specific spatial location (Gonzalez et al., 2009). RGB
components are important color features and are used by most of experts for studying minerals in thin
sections using polarizer microscopes.
4.2.2
HSI color space is an ideal space to develop image processing algorithm based on color descriptions that
are natural and intuitive for humans, who are the final developers and users of these algorithms (Gonzalez
et al., 2009). The hue of the color, as determined by wavelength, provides a measurement that directly
determines the color of the grain. Saturation is the amount of color that presents the degree to which a pure
color is diluted by white light (Thompson et al., 2001). Each pixel in thin section images has a unique
value of hue and saturation features, and these features are independent of the intensity of the microscopes
source light.
10
4.3. Clustering
Clustering is one of the most important problems in computer science (Prudent, Ennaji, 2005). It aims
toat discovering groups of similar patterns and to identifying meaningful structures or useful patterns in the
data. There has been a lot of research on clustering, and many algorithms such as k-means, Fuzzy k-means
and ART (Adaptive Resonance Theory) have been proposed in the literature (Xu and Wansch, 2005). The
proposed method is based on an incremental clustering algorithm, which will be explained in details in the
following subsections.
4.3.1 Incremental Clustering Algorithm
Anticipating the number of minerals in thin sections is too difficult. Thus, conventional clustering
algorithms such as k-means or ANNs are not appropriate for minerals classification. This is because in
these clustering algorithms, the number of clusters (minerals) must be determined before running the
algorithm. Thus, it is necessary to use an algorithm which does not need the number of clusters (minerals)
as the input parameter. Moreover, the algorithm must take new patterns and assign them into former
clusters or create a new cluster.
Since the ART algorithms have online incremental learning capabilities, the basic idea to design
proposed clustering algorithm is adopted from ART. In this paper, a slightly modified version of the
incremental clustering algorithm proposed by Sadri et al., (2006) is used for segmenting minerals in thin
sections. This algorithm can run faster than k-means because it scans (visits) all the data points just once to
find the centers of the clusters. This property of the algorithm makes it suitable for clustering of online
incoming data (Sadri et al., 2006). Incremental clustering algorithm proposed by Sadri et al. (2006)
includes only one parameter, as (minimum similarity threshold), that must be adjusted before running the
algorithm. The value of is between 0 and 1, which where 0 means there is no similarity and 1 means two
objects are completely similar. In our modified version of this algorithm, pixels of input image are read
one by one by the clustering algorithm, and based on Eq. 2, the similarity between each input pixel and the
center(s) of previously created cluster(s) is determined, and the input pixel will be included in the cluster(s)
11
such that its similarity is greater than . If there is no such cluster, a new cluster will be created for that
pixel.
Similarity
Where
0 Similarity 1
(2)
xn] and Y=[ y1, y2, ...., yn] are the color features of two different pixels of X, Y. The steps of clustering by
the proposed algorithm are presented here, and a full description of it based on a simple example is also
presented in Fig. 6.
1. Adjust the minimum similarity threshold ().
2. Cluster-Centers-list .
3. Read the next input pattern.
4. Find similar clusters centers to the input pattern in the cluster-center-list with a similarity greater
than .
If found: assign the pattern to those clusters, and adjust the centers of those clusters.
If not found: build a new cluster and insert the input pattern to cluster-center-list as a new
cluster center.
5. Repeat steps in 3-4 for all the input patterns.
Then
\
Then
Then
Then
Then
Then
Then
Figure 6. An example of the proposed incremental clustering for a simple image including six pixels.
4.3.2
In order to segment thin sections containing altered minerals, some reported clusters are noisy clusters with
random pixel which normally cannot provide necessary information about index minerals. For example,
the noisy cluster presented in (Fig.7-C) is one of the reported clusters by the proposed algorithm in this
study for segmentation of images shown in (Fig. 7-A and 7-B). As can be seen here, this cluster (Fig.7-C)
does not provide any important information about index minerals. Therefore, a median filter with 33
mask size is used for removing salt and pepper noises, and slightly smoothing the boundary of the clusters.
13
In order to eliminate those blobs which are not removed by median filter, a threshold equal to 150 (which
this threshold is equivalent to 0.5% of the input image sizes) pixels is also used. After this stage, other
remaining clusters will be considered as candidate index minerals.
Fig. 7. An example of the noisy cluster. A: input image in plane polarized light, B: input image in cross polarized light and C:
noisy cluster which have not important role in final segmentation map. By using a median filter a threshold equal to 150 pixels
the noisy cluster is omitted and not used for creating final segmentation map
important steps in the proposed algorithm because weathering and alteration effects create minor color
changing through minerals surfaces. Combining similar clusters with an overlap greater than 85% can
handle these kinds of color changing. It also produces more reliable results, which are closer to qualified
human expert judgment for minerals segmentation in thin sections containing altered minerals.
( )
( )
(
Clusters A and B
will be combined
Fig. 8. Combination of simple clusters A and B. n(A) and n(B) show the numbers of pixels in A and B clusters, respectively, and
AB shows their intersection.
Fig. 9. A: Bronzite in plane polarized light, B: Bronzite in cross polarized light, C and D: two clusters as a part of segmentation
15
map for A and B. The overlap of these two clusters is greater than 85%, and the proposed method combine them as one mineral
cluster.
5. Experimental Results
In order to evaluate the performance of the proposed algorithm in minerals segmentation, several
experiments were conducted on the images of thin sections from our database, which wasas explained in
Section 3. In the following subsections, first our evaluation method is explained. Then the details of
experiments for finding the best value of the parameter and the best combination of input features for the
proposed clustering algorithm are shown. Afterwards, the results of the experiments on thin sections with
and without altered minerals are presented, and in the end, results of the proposed method are compared
with similar methods in the literature.
5.1 Method of evaluation
The proposed segmentation algorithm produces a segmentation map for each thin section. Normally, the
number of detected clusters is more than the number of actual minerals existing in thin sections, and some
of the detected clusters belong to the same mineral. In order to evaluate the segmentation map or clusters,
an expert was asked to determine the actual existing minerals in test thin sections. At the output of the
proposed algorithm, all clusters which belong to the same mineral are aggregated; the number of pixels in
these clusters is summed; and the result is divided by the number of pixels of that mineral in the thin
section image determined by the expert. This gives the segmentation accuracy for that specific mineral.
This process is repeated for all minerals existed in the thin section. In addition, the average segmentation
accuracy of the proposed algorithm is calculated for each thin section. The results are reported in the
following subsections.
5.2 Choosing the optimum value of the parameter
In order to estimate the optimum similarity threshold parameter () in the proposed clustering algorithm,
four experiments have beenwere conducted using different input feature combinations. The details of
16
these four experiments are presented in following paragraphs. For these experiments, ten (10) thin sections
have been chosen including five thin sections without altered minerals (see Fig. 10) and five thin sections
containing altered minerals (see Fig. 11).
A1
A2
B3
B2
C1
C2
D1
D2
E1
E2
Fig. 10. Five thin sections without altered minerals are used for finding the optimum value of . A1 and A2 refer to Amphibole
index mineral in plane and cross polarized lights, B1 and B2 refer to Biotite index mineral, C1and C2 refer to Nosane index
mineral, D1 and D2 refer to Quartz index mineral and E1 and E2 refer to Sanidine index mineral.
A1
A2
B3
17
B2
C1
C2
D1
E1
D2
E2
Fig. 11. Five thin sections containing altered minerals are used for finding the optimum value of . A1 and A2 refer to Actinolite
index mineral in plane and cross polarized lights, B1 and B2 refer to Aegirin index mineral, C1and C2 refer to Amphibole index
mineral, D1 and D2 refer to Chlorite index mineral and E1 and E2 refer to Enstatite index mineral.
The first experiment was carried out using a 112 feature vector including RGB and HSI components in
plane and cross polarized lights for each pixel. The results of this experiment are given in Table 2. For
both types of thin sections, the =0.90 yields more accurate results. In Table 3, the results of the second
experiment using a 16 feature vector including RGB components in plane and cross polarized lights for
each pixel is presented.
2
Plane
Color _ component
2
Cross
more accurate results for this experiment. Furthermore, Table 4 indicates the results of third experiment
using 16 feature vector including HSI components in plane and cross polarized lights for each pixel.
Where =0.90 has more accurate results. The final experiment was conducted using a 16 feature vector
including a nonlinear combination of RGB and HSI color components in plane and cross polarized lights
based on Eq. 4. The results are presented in Table 5 where =0.9 yields to the most accurate results.
(3)
18
Table 2. The results of finding the best value of . A 112 vector including RGB and HSI components in plane and cross
polarized lights is established for each pixel. Both states of the thin sections have higher total accuracy when =0.90
Thin sections
Thin sections
without altered
containing
minerals
altered minerals
Accuracy when
Accuracy
Accuracy
Accuracy when
Accuracy
Accuracy
=0.80
when
when
=0.80
when
when
=0.85
=0.90
=0.85
=0.90
Amphibole
0.57
0.86
0.88
Actinolite
0.50
0.58
0.67
Biotite
0.60
0.80
0.90
Aegirin
0.50
0.83
0.92
Nosane
0.57
0.71
0.86
Amphibole
0.63
0.75
0.88
Quartz
0.57
0.86
1.00
Chlorite
0.54
0.85
0.85
Sanidine
1.00
1.00
1.00
Enstatite
0.57
0.71
0.71
Total
0.66
0.85
0.93
Total
0.55
0.74
0.81
accuracy
accuracy
Table 3. The results of finding the best value of . A 16 vector including RGB components in plane and cross polarized lights
is established for each pixel. Both states of the thin sections have higher total accuracy when =0.90
Thin sections
Thin sections
without
containing
altered
altered minerals
minerals
19
Accuracy
Accuracy
Accuracy
Accuracy when
Accuracy
Accuracy
when =0.80
when
when
=0.80
when
when
=0.85
=0.90
=0.85
=0.90
Amphibole
0.41
0.80
0.90
Actinolite
0.50
0.53
0.58
Biotite
0.52
0.71
0.83
Aegirin
0.48
0.75
0.85
Nosane
0.42
0.60
0.90
Amphibole
0.55
0.71
0.84
Quartz
0.51
0.79
1.00
Chlorite
0.43
0.72
0.78
Sanidine
1.00
1.00
1.00
Enstatite
0.49
0.57
0.65
Total
0.57
0.78
0.93
Total
0.49
0.66
0.74
accuracy
accuracy
Table 4. The results of finding the best value of . A 16 vector including HSI components in plane and cross polarized lights is
established for each pixel. Both states of the thin sections have higher total accuracy when =0.90
Thin sections
Thin sections
without
containing
altered
altered minerals
minerals
Accuracy
Accuracy
Accuracy
Accuracy when
Accuracy
Accuracy
when =0.80
when
when
=0.80
when
when
=0.85
=0.90
=0.85
=0.90
Amphibole
0.57
0.85
0.87
Actinolite
0.50
0.62
0.70
Biotite
0.67
0.80
0.86
Aegirin
0.58
0.79
0.85
Nosane
0.43
0.78
0.87
Amphibole
0.73
0.74
0.90
Quartz
0.57
0.80
1.00
Chlorite
0.62
0.83
0.86
Sanidine
0.72
0.85
0.91
Enstatite
0.41
0.65
0.80
Total
0.59
0.82
0.90
Total
0.57
0.73
0.82
accuracy
accuracy
20
Table 5. The results of finding the best value of . A 16 vector including nonlinear combination of plane and cross polarized
lights in RGB and HSI components (Eq. 3) is established for each pixel. Both states of the thin sections have higher total
accuracy when =0.90
Thin sections
Thin sections
without
containing
altered
altered minerals
minerals
Accuracy
Accuracy
Accuracy
Accuracy when
Accuracy
Accuracy
when =0.80
when
when
=0.80
when
when
=0.85
=0.90
=0.85
=0.90
Amphibole
0.51
0.87
1.00
Actinolite
0.45
0.55
0.80
Biotite
0.64
0.85
0.95
Aegirin
0.65
0.75
0.90
Nosane
0.70
0.77
0.90
Amphibole
0.71
0.80
0.92
Quartz
0.60
0.80
1.00
Chlorite
0.77
0.77
0.86
Sanidine
1.00
1.00
1.00
Enstatite
0.51
0.75
0.80
Total
0.69
0.86
0.97
Total
0.62
0.72
0.86
accuracy
accuracy
The overall results of these four experiments are shown in (Fig. 12). As shown in (Fig. 12), =0.90 and
nonlinear combination of plane and cross polarized lights in RGB and HSI color spaces as input features
yield the best results. This setup is fixed and used for final evaluation of the proposed algorithm.
21
Fig. 12. The accuracy of four different experiments of input feature vectors when =0.90. A 16 vector as input data set based
on nonlinear combination in plane and cross polarized lights for RGB and HSI components (Eq. 3) presents better results in
overall accuracy. Therefore, in the proposed algorithm =0.90 and nonlinear combination of plane and cross polarized lights as
input feature vector are fixed and used for minerals segmentation in thin sections.
22
A2
A1
B1
B3
B2
B4
B6
B5
Fig 13. The results of a thin section without altered minerals including Quartz index mineral (D1 and D2 in Fig .10). A1 and A2
refer to input data that assign to the clustering algorithm including nonlinear combination of the plane and cross polarized lights
in RGB and HSI components (Eq. 3), respectively. B1-B6 refer to the output of the clustering algorithm. Every black and white
image indicates a mineral cluster in the thin section images.
On the other hand, a thin section containing the altered minerals is used to evaluate the proposed algorithm
performance. As shown in (Fig. 14), a thin section containing the altered minerals with noisy pixels was
segmented by the proposed algorithm. It is important to note that the previous methods in the literature are
unable to segment the thin sections containing altered minerals such as the one shown in (Fig. 14). As seen
in Table 6, the overall accuracy of 92.15% and 85.24% has been obtained for thin sections without altered
minerals and thin sections containing altered minerals, respectively.
23
A1
A2
B1
B2
B3
B4
B5
B6
B7
B8
B9
B10
B11
Fig 14. The results of a thin section containing altered minerals including Aegirin index mineral (B1 and B2 in Fig .11). A1 and
A2 refer to input data that assign to the clustering algorithm, including nonlinear combination of the plane and cross polarized
lights RGB and HSI components (Eq. 3), respectively. B1-B11 refer to the output of the clustering algorithm. Every black and
white image indicates a mineral cluster in the thin section images.
Table 6. The final accuracy of the proposed method based on the database used.
Thin sections without altered minerals
43
122
24
92.15
85.24
Segmentation
process:
converting low
level pixel-based
information to
high level
minerals cluster
information
Fig 15. The conversion of low level pixel-based information into high-level cluster information minerals
In order to compare the results of the proposed method with the two previous methods as Zhou et al.
(2004-A) and Fueten and Mason (2007), image samples in these papers were used. It is important to note
that the proposed method needs plane polarized and cross polarized images from the thin sections;
however, in most previous papers including these two studies, cross polarized images were not used.
Hence, the proposed method was run only on plane polarized images in order to compare with these two
25
methods. The comparison of the intelligent segmentation algorithm used in this study with the method
based on edge detection and region growing presented by Zhou et al. (2004-A) is shown in Fig. 16. In
addition, the comparison of the proposed method with edge detection and artificial neural network (ANN)
algorithm proposed by Fueten and Mason (2007) is shown in Fig. 17. As seen in these two figures, the
results of the proposed method compares favorably with these two methods.
Fig. 16 Comparison of the proposed intelligent method with edge detection and ANN techniques (Zhou et al., 2004). A1: input
image and A2: result of Zhou et al. (2004) method. The other black and white images (B1-B16) are a final segmentation map
created by the proposed intelligent method.
26
Fig. 17 Comparison of the proposed intelligent method with edge detection and ANN techniques (Fueten and Mason, 2007).
A1: input image and A2: result of the method proposed by Fueten and Mason (2007). The other black and white images (B1B11) are final segmentation map created by the intelligent method used in this study.
color spaces which were used by incremental clustering algorithm. Although most of the thin sections
created in industry are those containing altered minerals, the current state of the art methods are not able to
accurately segment minerals in such thin sections. However, the proposed method is able to segment
minerals in both thin sections with and without altered minerals with highly accurate results. In addition, as
a full automatic method, the proposed algorithm can be used in geological purposes which require real
time segmentation map of thin sections. Experiments indicate that the proposed methods results
outperform the results of other similar methods used in other studies cited in the literature. This is
particularly true for thin sections containing altered minerals. We obtained the overall accuracy of 92.15%
and 85.24% for segmenting thin sections without altered minerals and those containing altered minerals,
respectively.
In future, we would like to enrich our features set using texture and shape features which are extracted
from the thin section images.
Acknowledgments
We would like to express our gratitude to mineralogy and petrography laboratory at University of
Birjand that had very close collaboration with us by providing the thin sections of minerals.
References
Arnould, X., Coster, M., Chermant, J.-L., Chermant, L., Chartier, T., Elmoataz, A. (2001) Segmentation and grain size of
ceramics. Image Analysis and Stereology 20, 131135.
Allard, B. and Benn, K. (1989) Shape preferred orientation analysis using digitized images on a microcomputer. Computers &
Geosciences 15(3), 441-448.
Barraud, J. (2006). The Use of Watershed Segmentation and GIS Software for Textural Analysis of Thin Sections, Journal of
Volcanology and Geothermal Research 154, 17-33.
Dyar M.D., Gunter, M.E. & Tasa, D. (2008), Mineralogy and Optical Mineralogy, Mineralogical Society of America, Chantilly,
Va. 708 p.
Fabbri, A. G. (1984) Image Processing of Geological Data. Van Nostrand Reinhold, New York, 224 pp.
28
Filho. I.M, Spina. T.V, Falcao. A.X, Vidal. A.C, (2013). Segmentation of sandstone thin section images with separation of
touching grains using optimum path fores operators. Computers and Geosciences, Vol. 57, pp146-157.
Fueten. F. (1997), A computer controlled rotating polarizer stage for the petrographic microscope. Computers & Geosciences
23, 203208.
Fueten, F. and Mason, J. (2007) An Artificial Neural Net Assisted Approach to Editing Edges in Petrographic Images Collected
with the Rotation Polarizer Stage. Computers & Geosciences 33, 1176-1188.
Gonzalez, R.C., Woods, R.E., Eddins. S.L. (2009) Digital Image Processing Using MATLAB.
Goodchild, J. Scott and Fueten, F. (1998) Edge Detection iv Petrographic Images Using The Rotation Polarizer Stage.
Computers & Geosciences 24(8), 745-751.
Heilbronner, R. (2000), Automatic Grain Boundary Detection and Grain Size Analysis Using Polarization Micrographs or
Orientation Images. Journal of Structural Geology 22, 969-981.
Izadi. H, Sadri. J, Mehran. N.A, (2013), "Intelligent Mineral Identification Using Clustering and Artificial Neural Networks
Techniques", Publihed in the Proceedings of The First Iranian Conference on Pattern Recognition and Image Analysis
(PRIA 2013), Birjand, Iran.
Lu, B., Cui, M., Liu, Q., Wang, Y. (2009) Automated Grain Boundary Detection Using the Level Set Method. Computers &
Geosciences 35, 267-275.
Obara, B. (2007), A New Algorithm Using Image Color System Transformation for Rock Grain Segmentation. Mineralogy and
Petrology 91, 271-285
Prudent, Y., Ennaji, A. (2005) A new learning Algorithm for Incremental Self-Organizing Maps. Proc. Of ESANN 2005 (13th
Europiean Symposium on Artificial Neural Network), Belgium, April.
Rashedi, E., Nezamabadi-pour, H. and Saryazdi, S. (2009) GSA: A Gravitational Search Algorithm. Information Sciences 179,
pp. 2232-2248.
Ross, B.J,, Fueten, F., Yashkir, D.Y. (2001), Automatic mineral identification using genetic programming, Machine vision and
applications, vol. 12, pp: 61-69.
Rui, X., Donald, W. (2005) Survey of Clustering Algorithms. IEEE Trans. On Neural Networks, Vol. 16, NO. 3, May.
Sadri, J., Ching. Y.S., Bui.T.D. (2006) A New Clustering Method for Improving Plasticity and Stability in Handwritten
Character Recognition Systems, 18th International Conference on Pattern Recognition.
Simigian, S. and Starkey, J. (1989) IMAGE: modied for use on a microcomputer based system. Computers & Geosciences
15(3), 237-254.
29
Thompson, S., Fueten, F., Bochus, D. (2001) Mineral Identification Using Artificial Neural Networks and the Rotation Polarizer
Stage. Computers & Geosciences 27, 1081-1089.
Van den berg, E. H., Meesters, A. G. C. A., Kenter, J. A. M., Schlager, W. (2002) Automated Separation of Touching Grains in
Digital Image of Thin Sectins 28, 179-190.
Yesiloglu-Gultekin, N., Keceli, A.S., Sezar, E.A., Can, A.B., Gokceoglu,C., Bayhan, H. (2012), A Computer Program
(TSecSoft) to Determine Mineral Percentages Using Photographs Obtained from Thin Sections. Computers &
Geosciences46, 310-316.
Zhou,Y., Starkey, J., Mansinha, L. (2004A), Identification of Mineral Grains in a Petrographic Thin Section Using Phi- and
Max-Images, Mathematical Geology 36(7), 781-801.
Zhou,Y., Starkey, J., Mansinha, L. (2004B), Segmentation of Petrographic Images by Integrating Edge Detection and Region
Growing, Computers and Geosciences 30, 817-831.
Highlights
30