Vous êtes sur la page 1sur 5

Earth Rock Classification using Neural Network

Pratik V. Patel1, Dr. Himanshu S. Mazumdar2


Research Scholar, E-mail: patelpratik512@gmail.com
2
Professor & Head, E-mail: hsmazumdar@hotmail.com
Research and Development Center, Dharmsinh Desai University, Nadiad, Gujarat, India.
1

sensors at different magnifications. The color, texture


and grain are important indicators of its chemical and
structural properties of rocks. These parameters are
used for pattern recognition of the rocks [2][3].The
combined analysis of these results increases the
accuracy and reliability of prediction of rock type.
This paper describes the algorithms developed for
these three image parameters.
The work presented here uses earth rock samples to
verify the proposed classification algorithm. Five such
samples of common river rocks are shown in
Figure1.The algorithm is expected to work for in-situ
analysis of moon rocks using rover. Three sample
moon rock images [4] with corresponding microscopic
images are shown in Figure 2.
Sixty earth rock samples are collected from
MahiSagar river bed and cataloged in image gallery.
Six microscopic grain images with magnification of
100 are taken of each rock sample. These 360 grain
sample images were used for classification of 60 rock
samples.

Abstract - Objective of this paper is to present a


method for Rock Classification system using
microscopic imaging of surface parameters. The
development in the area of computational
intelligence has opened new opportunities in the area
of image processing. In this paper we demonstrate a
similar example of earth rock classification. This
consists of feature extraction using wavelet based
data compression and neural net based feature
classification. Classification accuracy is further
improved using multi-parameter analysis of different
surface parameters. Rock surface parameters used in
this work are color, texture and grain. The
combined signature extracted from each of these
parameters is used to identify the rock type. The
project is developed under the Planetary Exploration
Technology Research project.
Keywords color, texture, grain, image processing,
wavelet based data compression, feature extraction,
neural net based classification.

1. Introduction
D.D. University has under taken a project AstroMaterial Analysis under the Planex program of
Physical Research Laboratory, ISRO. Rocks are an
important category of astro-materials. Rocks are the
major composition of most planets like earth, moon
and mars. The rocks are indication of evolution of a
planet and its composition. Analysis and interpretation
of earth rocks are mostly applicable to the rocks of
other planets and meteorites. Analysis of lunar rock
sample has greatly helped in understanding of its
mineral, chemical composition and age of associate
important events. Important activities like evolution of
a planet crust, the timing of critical events like
meteorite bombardment and volcanism are reflected in
the property of its surface rocks.
Classification of rocks using multi-sensor pattern
recognition algorithms provides a simple and efficient
technique in planetary exploration programs [1].We
propose to measure five different parameters of each
rock sample and classify for rock identification. These
parameters are color, texture, grain, scratch and sound.
The first three parameters are acquired using imaging

Fig 1. Top row shows five selected earth rock samples


with corresponding magnified grain images at bottom
row. The grain images are acquired with 100x digital
microscope.

Fig 2. Top row shows the sample moon rock images


[4] of Mare basalt, Impact breccias and Plutonic rock
(Anorthosite) along with the corresponding grain
images in bottom row.

The color, texture and grain parameters are derived


from these grain images using different spatial
sampling intervals. Figure 3 shows the block diagram
of the proposed system to measure and analyze color,
texture and grain parameter of rock samples. The brief
description of the analysis is given belowColor analysis: Most popular image format of digital
camera is 24 bit RGB color, which contains intensity
of red, green and blue color components in 8 bit gray
scale. These values are dependent on illumination
level and reflectance of the object. George Paschos [7]
compared and shown in his work that HSV color space
performs better than L*a*b and RGB .To extract the
color parameter independent of external illumination,
the RGB space is converted to Hue, Saturation and
intensity Value (HSV) parameters, where Hue
represents the color component independent of color
saturation and illumination intensity. The histogram of
hue values are computed using random samples taken
from entire image space. Hue histogram is then
transformed using Daubechies wavelet. Transformed
data is used as input to a Multilayer Feed Forward
type Neural Network. The network is trained using
rock type number as desired output encoded using
binary walking code [6]. The result of analysis shows
95% of average accuracy in prediction of training
sample and 88% of average accuracy in prediction of
new untrained sample as shown in Table 1.
Texture analysis: Rocks are commonly identified by
their texture. Mari Partio [10] presents gray level cooccurrence matrix (GLCM) to evaluate rock images
based on texture. In the work of Leena Lepisto [8],
they shown average 75% result for color based rock
texture classification. Here the sampling method used
is color pair co-occurrence population. The most
occurring hue color pairs are used for the analysis. The
color pair is selected randomly within range of 60
pixels (+/-30) distance. The distance 60 is chosen to
average 30x30 pixels area to represent texture. The
result of analysis shows 98% of average accuracy in
prediction of training sample and 84% of average
accuracy in prediction of new untrained sample as
shown in Table1.
Grain analysis: This is similar to the texture analysis
except the hue color pair is selected randomly within
range of 6 pixel (+/-3) distance. The distance 6 is
selected to average 3x3 pixels area to represent grain.
The result of analysis shows 98% of average accuracy
in prediction of training sample and 86% of average
accuracy in prediction of new untrained sample as
shown in Table1.

Fig 3. Block diagram of proposed system.

Fig 4. Shows 60 Rock samples collected from


Mahisagar River. Ten randomly selected rocks with
circle marked are used for analysis.

Fig 5. Screenshot of application software developed


for rock classification.

2. Proposed Method:
Three parameters of rock samples, color, texture and
grain are analyzed using the method described in
section 1. All grain images used for this experiment
are in JPEG format of resolution 1600X1200 in RGB
color. Data is collected by sampling random pixels in
the image area. Sampling methods for each of
parameter is described in next section.

identify rock type. Configuration of each NN used is


three layer feed-forward type connections and consists
of input, hidden and output layers. The NN for color
Parameter uses 180 input neurons. The other two NN
for texture and grain parameters are similar and uses
235 input neurons. All inputs are normalized before
applying to NN. All the three NN is configured using
20 hidden neurons and 10 output neurons. Each output
corresponds to a rock type. Only one output is high at
any time. Network is trained by computing the output
error with respect to the desired output [16]. Numbers
of hidden neurons are manually optimized for speed
and accuracy. The configuration of NN used for
classification is shown in Figure 6. The Input, hidden
and output layers are Yi, Yj and Yk respectively.
Desired output for training is dk. Xn and Yn are inputs
and outputs of neurons. Output of NN is computed
using equation 1, 2, 3 and 4. Network is trained using
error function given in equation 5. Error Sensitivity
E/ W of weights is calculated to reinforce the
weights as shown in equation 6 and 7.

2.1 Sampling Method:


For color analysis each sample consists of single pixel
value. For texture and grain analysis each sample
consists of a pair of pixels. For texture sample the
paired data is sampled between distance ranges of
random +/- 30 whereas the grain sample uses distance
range of random +/- 3. Total number of random
samples used is in the range of 30,000 to 19, 20,000
depending on the accuracy and speed of recognition.
These RGB samples are converted to HSV color
format. In case of texture and grain analysis only most
occurring pairs are used to restrict the input size
whereas in case of color analysis most occurring 180
hue values are used from full range of 360. The hue or
hue pair value is used to calculate normalized
histogram. The further processing consist of wavelet
transform and neural network classification as
described in section 2.2 and 2.3 below.

2.2 Wavelet Transform:


Wavelets are used to convolve shape and texture with
location information of image [11][12]. All three
features in our application is subjected to Daubechies
[13] wavelet. Daubechies D4 scaling function (ai) and
wavelet function (ci) are shown in equation-(1) and (2)
respectively. The scaling and wavelet functions are
calculated by taking the inner product of the
coefficients and four data values. A segment of
transformed values are selected and used for
classification. Different filter profile may be used to
select appropriate segment of transformed values to
optimize the accuracy and speed of classification.
These selected transformed values are further
processed using multilayer Neural Network.

Fig 6. Shows a three layer Neural Network with Yi as


inputs, Yk as outputs.

ai =h0s2i + h1s2i+1 + h2s2i+2 + h3s2i+3


--------- (1)
Where h0 = (1+3)/42, h1 = (3+3)/42, h2 = (3-3)/42
and h3 = (1-3)/42.
ci = g0s2i + g1s2i+1 + g2s2i+2 + g3s2i+3 --------- (2)
Where g0=h3, g1=-h2, g2=h1 and g3=-h0

2.3 Neural Network Architecture :

Xj = i=0Yi * Wji
Yj = 1/ (1+e-xj)
Xk = j=0Yj * Wkj
Yk = 1/ (1+e-xk)

-------------------------------------- (1)
-------------------------------------- (2)
-------------------------------------- (3)
-------------------------------------- (4)

E = 0.5*(Yk-dk) 2

--------------------------------------

(5)

Wkj = Wkj + eta * (Yk-dk) * Yk * (1-Yk) * Yj ----- (6)


Wji = Wji + eta * k=0{(Yk-dk) * Yk * (1-Yk)}
* Yj * (1-Yj) * Yi
---------- (7)

Three sets of wavelet transformed values are


separately classified using multilayer Neural Network
(NN) [14] with back-propagation training algorithm to

On completion of training network weights are stored


for online prediction of rock samples using equation 1,
2, 3 and4.

type the signature mismatch is greater than a


predefined value and is consider as insufficient
sample. The algorithm messages Unable to Find
Rock Type and this type of error are not included in
result of Table1.
The test set and the training set used in the simulations
are mutually exclusive.

3. Experimental Setup:
For the experiment 10 different rocks were randomly
selected from 60 rock types described in section I.
These rocks were photographed using digital
microscope camera with two different magnifications.
The low magnification is used to capture full image of
the rock and magnification 100x is used for grain
imaging. Test rock dataset consists of 10 types of rock
images as shown in Figure 4. Fifty grain samples are
taken from each rock. A total of 500 grain images are
used for analysis. Figure 7 shows experimental setup
of proposed work.
The texture analysis use larger segmentation of the
grain image to simulate the low magnification imaging
and grain analysis use smaller segmentation of the
grain image to simulate the high magnification
imaging.

Table1. Result of earth rock classification


Average accuracy
percentage of
No. of
No. of
10 random trails
Feature Training Testing Untrained Samples
images
images
test
used for
Samples
training
(%)
(%)
490
10
88
95
Color
10
84
98
Texture 490
10
86
98
Grain 490
Table 2. Average % of 10 sets of untrained
testing sample
Feature
Set No.
Set 1
Set 2
Set 3
Set 4
Set 5
Set 6
Set 7
Set 8
Set 9
Set 10
Average %
of all set

Fig 7. Experimental setup of rock classification

4. Experimental Results :
Result presented here are from 10 rock types as
described in section 3. These 500 image dataset is used
for the training and testing of all three parameter.
From this dataset other two datasets are prepared by
randomly selecting 10 images for testing and 490 for
training of texture and grain analysis. Each image
contains 19, 20,000 pixels (1600x1200). For each test
experiment we use
1, 00,000 color or color-pair
samples. From each of 10 test images, 10 random
trails of 1, 00,000 random samples are taken to
evaluate the algorithm. Table 2 shows the result of 10
such sets of data. Table 1 shows the overall result of
rock classification with trained and new untrained
samples. There are two types of errors, in first type the
algorithm predicts wrong rock group which we
consider as error in the result of table1, in the second

Color
(%)

Texture
(%)

Grain
(%)

100
89
83
100
100
60
80
90
88
90
88

82
90
85
64
100
60
93
90
89
90
84

89
100
76
81
90
85
85
90
89
81
86

5. Conclusion:
A Neural Net based approach is presented here that
allows user to classify the rocks using microscopic
image of target rock surface. The technique used here
classifies the rock type with 95% of average accuracy

Gabor Filtering , Proceedings of the 12th International


onference on Image Analysis and Processing (ICIAP03) 07695-1948-2/03 $17.00 2003 IEEE.
[10]Mari Partio, Bogdan Cramariuc, Moncef Gabbouj, Ari
Visa, Rock Texture Retrieval Using Gray Level CoOccurrence Matrix, Tampere University of Technology,
P.O. Box 553, FIN-33101 Tampere, Finland.
[11]Dipankar Hazra, Texture Recognition with combined
GLCM, Wavelet and Rotated Wavelet Features.
International Journal of Computer and Electrical
Engineering, Vol.3, No.1, February, 2011
[12]Amara Graps An Introduction to Wavelets, 1995
Institute of Electrical and Electronics Engineers. published
by the IEEE Computer Society, 10662 Los Vaqueros Circle,
Los Alamitos, CA 90720, USA,TEL +1-714-821-8380, FAX
+1-714-821-4010.
[13]Stephane Mallat, A Wavelet Tour of Signal
Processing. Academic Press, pp. 98-405, 1999.
[14]Mehdi Lotfi, Ali Solimani, Aras Dargazany, Hooman
Afzal, Mojtaba Bandarabadi, Combining Wavelet
Transforms and Neural Networks for Image Classification.
41st Southeastern Symposium on System Theory University
of Tennessee Space Institute Tullahoma, TN, USA, March
15-17, 2009.
[15]Heather Dunlop, Automatic Rock Detection and
Classification in Natural Scenes. Robotics Institute
Carnegie Mellon University Pittsburgh, Pennsylvania 15213
Carnegie Mellon University, Aug 2006.
[16]Himanshu S. Mazumdar & Leena P. Rawal, "A Neural
Network Tool Box using C++ ", in CSI Communications,
August, pp. 15-23, Bombay, 1995.
[17]R. A. Yingst1, B. A. Cohen2, B.M. Hynek3, J.B.
Johnson4, M.E. Schmidt5 and C.M. Schrader2, ScienceDriven Strategies For Semi-Autonomous Rovers On The
Moon: Field Test At An Ice-Bearing Regolith Analog. 43rd
Lunar and Planetary Science Conference (2012) USA.
[18]Martha S. Gilmore, Rebecca Castao, Tobias Mann,
Robert C. Anderson, Eric D. Mjolsness, Roberto Manduchi,
R. Stephen Saunders. Strategies for autonomous rovers at
Mars. Journal of Geophysical Research Planets, USA,
April 18, 2000.
[19]Lunar Rocks, chapter 6 of Lunar Sourcebook,
Lunar and planetary institute, Houston, Texas Cambridge
University Press 1991.USA.

of trained samples and 86% average accuracy of new


untrained samples which is greater than the Leena
Lepistos [8] average 75% result of rock image
classification. Results obtained with wavelet and
neural net based algorithm is encouraging and could
be used in similar surface classification application.
Larger training and test sets of rock images could
provide more reliable and accurate results.

Acknowledgment
This work is carried out under the project Planetary
Exploration Technology Research, funded by
PLANEX program of Physical Research Laboratory,
ISRO, and India.

References:
[1] Himashu Mazumdar, Agnel Amodia, A Pattern
Recognition Framework for Embedded Sensor Electronics
International Journal of Recent Technology and Engineering
(IJRTE) ISSN: 2277-3878, Volume-I, Issue-I, April 2012.
[2] Naresh Singh, T. N. Singh, Avyaktanand Tiwary, Kripa
M. Sarkar, Textural identification of basaltic rock mass
using image processing and neural network, Comput Geosci
(2010) 14:301310 DOI 10.1007/s10596-009-9154-x
[3] Laercio Brito Gonalves and Fabiana Rodrigues Leta,
Macroscopic Rock Texture Image Classification Using a
Hierarchical Neuro-Fuzzy Class Method Hindawi
Publishing Corporation Mathematical Problems in
Engineering Volume 2010, Article ID 163635, 23 pages
doi:10.1155/2010/163635
[4] Moon
rock
sample
collection
site,
http://curator.jsc.nasa.gov/lunar/samplecatalog/index.cfm#,
http://curator.jsc.nasa.gov/lunar/compendium.cfm
[5] Binary
Walking
Code
technique,
http://www.springerreference.com/docs/html/chapterdbid/28
386.html.
[6] Pengyu Liu, Kebin Jia and Zhuozheng Wang., An
Effective Image Retrieval Method Based on Color and
Texture Combined Features. Beijing University of
Technology, Beijing China.
[7] Paschos, G.; , "Perceptually uniform color spaces for
color texture analysis: an empirical evaluation," Image
Processing, IEEE Transactions on , vol.10, no.6, pp.932937,Jun 2001doi: 10.1109/83.923289
[8] Leena Lepist, Iivari Kunttu,Ari Visa, Rock image
classification using color features in Gabor space", Journal
of Electronic Imaging, 2005/14(4)/040503.
[9] Leena Lepist, Iivari Kunttu1, Jorma Autio, and Ari Visa,
Classification Method for Colored Natural Textures Using

Vous aimerez peut-être aussi