Real-Time Imaging 8, 1–9 (2002)

doi:10.1006/rtim.2000.0255, available online at http://www.idealibrary.com on

A New Approach to Identify Big Rocks

with Applications to the Mining Industry

D
etection of big rocks is an important, even critical, problem in the mining industry due to
the risk of machine blockage causing high costs. This paper presents a computer-vision-
based method to detect big rocks in a real mining industry. Our system, based on a
mixture of image processing techniques and neural networks, works as follows: once the image is
taken, a pre-processing step is performed, filtering the image and extracting a set of candidate
rocks. Then a neural network processes the candidate rocks to ensure correct detection. A
tracking algorithm is then applied to avoid false detection due to rock grouping. Using
geometrical information, it is possible to estimate the real dimensions of the rocks. Our computer
vision system satisfies time constraints imposed by the industry to work in real time and is
currently operating. The algorithm presented is independent of the rock’s shape. Results obtained
during nine months of unsupervized work are provided, showing that our system is able to work
under different light conditions and is robust enough to face real work conditions.
# 2002 Elsevier Science Ltd.

Enrique Cabello*, M. Araceli Sánchezw and Javier Delgadoz

*Universidad Rey Juan Carlos, ESCET, C/Tulipán s/n, 28933 Móstoles, Madrid, Spain.
E-mail: ecabello@escet.urjc.es
w
Universidad de Salamanca, Departamento de Informática y Automática,
Plaza de la Merced s/n, 37008 Salamanca, Spain
z ENUSA, Crta. Ciudad Rodrigo-Lumbrales, Km 7, 37500 Ciudad Rodrigo, Salamanca, Spain

Introduction allows for the implementation of a low-cost vision

system. This environment is very hostile for the opera-
ENUSA (Uranium National Company of Spain) owns tion of such a system mainly due to the following facts:
an opencast mine near Salamanca (Spain). A problem in
this kind of industry is the estimation of the size of K Dust is in the air, and may totally or partially occlude
rocks. The problem arises because big rocks can block the visibility of the scene, especially when a truck
crushers, vibrator feeders or conveyor belts. Such unloads rocks in the bin.
situations lead to high costs and may be dangerous for K The area covered by the camera has to be illuminated
human operators. The traditional approach is to employ constantly to avoid loss of information. On the
visual inspection to identify big rocks, but a more contrary, in our system the sun may illuminate the
efficient solution is an automatic inspection using digital chutes and feeders modifying the quantity of light.
image processing. K Rocks have no uniform colours. In our case the rocks
are sometimes wet and therefore almost black, which
makes it very hard to extract their shape.
A computer vision system could help the worker,
increasing security and saving money for the company. In this paper, we present a new approach to identifying
The dramatic reduction in computer and camera costs big rocks, based on a mixture of image processing

1077-2014/02/010001+9 $35.00/0 r 2002 Elsevier Science Ltd.

2 E. CABELLO ETAL.

techniques and neural networks. Based on this ap- metric distribution of rocks on a conveyor belt is
proach, a real-time application has been developed. We computed with excellent results under the assumption of
have established an efficient algorithm to identify big controlled light.
rocks for a mining company. Excellent results have been
achieved in unsupervised operations (more than 70%
correct identification during a nine month period),
Proposed Rock Detector
working in real time and in a hard environment.
System overview
The structure of the rest of the paper is as follows.
The next section provides a short overview of previous
In the Uranium mine of our study there are two chutes
research. The section following is devoted to the
and two conveyor belts. One PC, two image acquisition
description of our system and the detailed explanation
cards and two black and white CCD cameras make up
of the algorithms implemented. Then, we present the
our vision system. Each camera is plugged into an
numerical results and a discussion. We end with
acquisition card and is focused on one chute. The
conclusions in the last section.
resolution for the image is set at 640  480 pixels with
256 gray levels. The software switches between the
cards, taking and processing alternatively one image
Previous Research
from each camera. A schematic vision of the system is
shown in Figure 1, showing both chutes and both
Mining industries are an open field for the use of
cameras. It is shown that two separated images are
computer vision systems in a large number of activities
analysed, each one obtained by a camera and containing
[1]. In particular, computer vision techniques could be
one chute.
employed to improve the quality of the mineral and to
avoid risky operations. The estimation of a rock’s size
Figure 2 shows a chute during operation. To the right
on a conveyor belt is useful to avoid blocking situations
of the image two metallic chains can be seen whose
or to determine the granulometric distribution of the
purpose is to avoid the rocks getting on top of one
rocks. Different algorithms have been considered in
another. Also, the chains sweep the rocks, forcing the
recent years to detect a rock on a conveyor belt as
biggest dimension of the rock into the plane of vision.
described below.
Dimensions of each chute are 2.5 m width and 6.5 m
long. Speed of the rocks in the chute is difficult to
An image-classification algorithm is presented by
measure and, in fact, no such measures were available at
Wang and Stephansson [2] to estimate the character-
the beginning of the work. Our estimations correspond
istics of rock fragments (size distribution and shape),
from 0.5 to 0.8 m/s, depending on the size and number
but its results depend on the quality of the image. A
of rocks, amount of dust in the chute and other
technique based on the shadows surrounding a rock is
conditions. The system achieves a processing rate of 8
presented by Wu et al. [3] but it is quite sensitive to light
image/s (4 image/s is per chute), enough to control both
conditions and rock texture. A multiresolution system is
chutes.
presented by Crida et al. [4,5] based on 12 features and
on the knowledge of rock characteristics. Experimental
thresholds for each feature have to be chosen.

Artificial Neural Networks have been used by Crida

and De Jager [6] and Fernandez et al. [7]. Crida and De
Jager [6] considered a feed-forward neural network
trained with a backpropagation algorithm to detect
rocks. The preprocessing step uses contrast enhance-
ment and edge enhancement. No numerical results are
provided but the authors stated that artificial neural
networks alone are not enough to recognize a rock.
Fernandez et al. [7] used a retina and a coding technique
to allow an artificial neural network to estimate the
granulometric distribution of an image. The granulo- Figure 1. Schematic vision of the system.
 A NEW APPROACH TO IDENTIFY BIG ROCKS WITH APPLICATIONS TO THE MINING INDUSTRY 3

Figure 2. The chute during operational hours.

Due to the dimensions and location of the chutes, it is

unfeasible to efficiently control light sources. Both
chutes are placed in a building near one to another,
and the sun illuminates them in different ways during Figure 3. Scheme of the algorithm.
the day. To control illumination the proposed solution
will cover the chutes and allow only controlled
illumination. This solution was discarded by the chute. Figure 3 shows an overview of the decomposition
company due to its economic impact and the fact that of the algorithm. Each phase is later described in detail,
the building is used for another services. The accepted with an example.
solution consists of additional points of light to obtain a
more uniform illumination of both chutes. Four focuses Different algorithms were considered (different edge
by chute were placed to illuminate it. As electric power is detectors and segmentation schemes among others), but
cheaper at night most of the work is performed during their results were not suited to our project. For example,
night hours and in this case the focuses are an adequate edge detectors offer us a big set of small segments
light source. During both day and night, the focuses without any order because the rock’s shape is irregular,
provide adequate illumination, except towards late having a lot of corners. The segmentation algorithms
afternoon as the sun may illuminate the chutes and considered either led to a large number of small regions
saturate the cameras. This is an inherent obstacle to the or were very time consuming. Wet rocks are almost
actual configuration of the system. Therefore, even black and segmentation algorithms merge rocks with
illumination is not a serious problem as we have dust.
implemented an economic and effective control, allow-
ing the system to operate during most part of the day
and all the night. Image acquisition

Algorithm description In our set-up the camera is firmly placed, so the chute is
always shown in the same position. A region of interest
The algorithm has been structured in three phases is defined containing the chute, and only this region of
depending on the treatment applied to the images: interest will be processed. Those parts of the initial
acquisition, preprocessing (to obtain a set of candidate image outside the region of interest are eliminated, so
rocks in the image) and processing (using an artificial processing time is reduced without any loss of informa-
neural network and a tracking algorithm to ensure tion (see Figure 4).
correct detection). Once the preprocessing step is
finished, we have a set of candidate regions in the Processing rate (4 frame/s per chute) is almost
image. The processing step is initiated with the set of independent of the number of rocks in the image but
candidate regions and the initial gray level image, the depends on the number of pixels in the region of
result will be the correct detection of big rocks in the interest.
4 E. CABELLO ETAL.

Figure 4. Region of interest in the image. Figure 5. Equalized image.

Image preprocessing Step 2

This step deals with the softness of the image,
In this phase, quick algorithms are applied to obtain eliminating undesirable effects that could be present in
information about the regions in the image that may a digital image. A low-pass filter is used to eliminate
contain rocks. Some of the algorithms considered are high frequency components (see Figure 6). This high
standard (histogram equalization and low-pass filter [8]), frequency information represents high pixel-to-pixel
but others have been tailored to our problem (gray level variations associated with noise or with light reflection.
threshold and ‘‘separator’’ algorithms). Low frequency information represents the shape of the
rocks and is preserved. This step is implemented by the
Preprocessing is a critical step in which the input is a convolution of the image with a filter, and the selected
gray-level image and the final output is a set of windows kernel was
in the image that may contain a rock (a set of 2 3
1 1 1
‘‘candidate rocks’’). The pre-processing phase has to 1 4
1 2 1 5;
deal with all the problems related to gray-level images 10
1 1 1
and hence a robust algorithm has to be built. This
algorithm is split into several steps. Each one represents
the application of a process in order to isolate the rocks
in the image.

Step 1
The histogram is equalized [8,9] to obtain a uniform
distribution in the gray levels along the image pixels,
consequently the contrast is increased to achieve better
differentiation of the rocks (see Figure 5). This step tries
to mitigate the changes in the light conditions under
which the system can work. In images of wet rocks an
overall brightening is achieved. In general, spreading out
the brightness values makes small variations more
evident. However, equalization does not work well
when both sunlight and shadows are present, but in our
conditions this is a very singular case. Usually we have
the whole image in sunlight or in shadow, so this step
offers acceptable results. Figure 6. Low-pass filtered image.
 A NEW APPROACH TO IDENTIFY BIG ROCKS WITH APPLICATIONS TO THE MINING INDUSTRY 5

but the following were also considered: In our case, the direct thresholding of the image is not
2 3 able to produce a complete representation of rocks:
2 3 2 3 1 2 4 2 1 complex situations like shadows, overlapping and
1 2 1 1 1 1 62 4 6 4 27 touching rocks make this process very unreliable.
1 4 1 6 7
2 4 2 5; 4 1 1 1 5and6 64 6 8 6 477: Therefore a two-level algorithm is proposed. Our
16 9 42 4
1 2 1 1 1 1 6 4 25 algorithm is based on Tsai’s method (see Appendix)
1 2 4 2 1 and acts as follows. Using the histogram of the region of
interest, a first optimum threshold is computed using the
first three moments. Then, all the gray levels over the
Step 3 threshold are eliminated. With the remaining histogram
A two-level threshold process is then performed (see a second threshold is computed using the same algor-
Figure 7) in order to isolate the rocks present in the ithm. The final threshold is the mean between these two
image. Rocks are marked as white regions in a black thresholds. This two-level algorithm is capable of
background. To automatically compute the threshold, dealing with the noise in the image and performing well
the algorithms developed by Tsai [10], Otsu [11] and the with wet or dry rocks. Its computing time is reduced in
mean and median algorithm described by Jain [12] were comparison with that of the others methods considered.
tested. None of them offered good results for our
images. Step 4
A set of transforms is applied to the last image to
Several authors [13–15] have undertaken comparisons achieve a complete separation between white regions.
of the various thresholding methods, although a An algorithm based on erosion [16] is implemented (see
common difficulty has been to find an adequate metric Figure 8). The addition of this morphological-based step
for determining the effectiveness of each. Sahoo et al. offers better results in rock isolation, eliminating small
[13] considered the uniformity and shape of the white regions. The algorithm works as follows: if one of
thresholded objects to determine the success of several the eight neighbors of a pixel is black, this pixel will be
methods and concluded that Otsu’s algorithm was better black. This is applied eight times to give us smaller and
(although Tsai’s method was not considered). In this better-separated white regions. This number is set accor-
paper, Tsai compared their algorithm with Otsu’s ding to the size of the image and considering the
method and concluded that both produced similar minimum size that has to be detected. At the end, small
results but Otsu’s method performs poorly in the junctions between white regions are removed. Then, a
presence of noise and shadows. Another advantage is ‘‘separator’’ algorithm is applied, removing white reg-
that Tsai’s method is significantly faster. ions with few pixels both in vertical and in horizontal.

Figure 7. Thresholded image. Figure 8. Eroded Image.

6 E. CABELLO ETAL.

Erosion has to be applied together with a dilatation

operator that brings back the initial contours of the
rocks with minimal error. This dilatation takes place
during the next step, so the initial length of the region is
preserved.

Image processing

Once the erosion operator is applied, the system assumes

that each white element in the image is a rock. Each
white region is inscribed in a rectangle, whose size is
given by the dilatation operator (see Figure 9). This
region in the image, supposedly containing a rock, is
called a ‘‘candidate rock’’. Only rectangles with a size
bigger than a threshold are considered later by a neural
network, The size of the rectangle is an estimation of the Figure 10. Artificial neural network results.
real size of the rock. As the distance from the camera to
the chute is known and the focus is constant, a simple
geometric transformation gives us an estimation of the
real dimension of the rock based on the size of the In order to increase the accuracy of the results we
rectangle. have implemented the following tracking method: once
a region is identified as a rock, it is tracked over a
Artificial neural networks are a well-known classify- sequence of 10 images (approximately equivalent to 2.5
ing technique and have repeatedly shown their potential s). If it is detected in at least half the images of the
as a tool for classification purposes [17–19]. A neural sequence, it is considered as a blocking rock. To
network decides if a candidate rock can actually be consider that a big rock is the same in two frames, the
classified as a rock or not: the input is a gray-level image size of the region has to be similar and the displacement
(a candidate rock) and the output is its classification as a has to be performed according to the drooping direction
rock or not. Artificial neural networks are very useful to of the chute. (Rock droops are always in the same
detect if two or more rocks compose a candidate rock. direction for one chute). This method avoids false
detection due to rock crowding, since they will not
More details are given in the next section. Note that persist over several images. Once a blocking rock is
the neural network only analyzes regions whose size is detected, the operator is informed and the chute is
bigger than a threshold. The results in the image under stopped.
consideration are shown in Figure 10.

Numerical Results and Discussion

Two types of artificial neural networks were compared

in our study: three layers feed-forward trained with
back-propagation algorithm (FANN) and Radial Basis
Functions (RBF). The topology used was the same for
both networks (see Figure 11): an input layer with 400
(20  20) neurons, a hidden layer with 20 neurons and
an output layer with two neurons. The input of the
neural network is a ‘‘candidate rock’’, reduced to
20  20 pixels and pre-processed. A hidden layer of 20
neurons was considered after testing several numbers.
The activation of one neuron in the output means that
the pattern (the image) is a rock and the activation of
the other neuron in the output means that it is not a
Figure 9. Detection of possible rocks. rock.
 A NEW APPROACH TO IDENTIFY BIG ROCKS WITH APPLICATIONS TO THE MINING INDUSTRY 7

Table 2. Extended tests with softened images (test set)

% Of correct identification
FANN 92.0
RBF 94.1

used in the system is a RBF with softened images, due to

Figure 11. Scheme of the artificial neural network. its better performance.

Similar results have been found in [7] (88% correct

Several dimensional reduction techniques on the input detection with a Multilayer Preceptron and 94.8% with
image have been proposed in the bibliography; we could a MLP+RBF) but the database was formed only by 43
mention Principal Components Analysis (PCA) [20], images and 70% overall correct detection was found.
Discrete Cosine Transformation (DCT) [21], Most We could note that Crida [6] does not offer any
Discriminant Features (MDF) [21] or Gabor filters numerical result. Our algorithm has been tested
[22]. None of them has been considered in the rock intensively during nine months of unsupervised work
detection problem (see [6,7]) because these techniques and in several working conditions. We think that these
result in long processing times, and speed is one of the results demonstrate the validity of our approach.
constraints of the system. In our work, we have
preferred to use only the data stored in memory (even Initial in-place tests were performed to gain an in-
if we have to increase the number of neurons) than to depth understanding of the problem under real opera-
perform more processes. To obtain the feature extractor tion and to familiarize the worker with the system. Once
we take into account the results of the different the system was in place, we studied its behavior and
processes applied to the initial image. Table 1 shows some feedback information was provided about condi-
the results considering gray-level normalization, an tions under which no detection or incorrect detection
equalization step before normalization, and a low-pass occurred.
filter before normalization of the input image.
Figure 12 shows the results during nine months (from
At the beginning only 200 images were available, so a January to September, 1999) of unsupervised work with
‘‘cross-validation’’ technique was considered to train the the final version of the system. ‘‘Correct detection’’
neural networks. An initial set of 170 training and 30 means that a blocking rock was detected by the system.
testing images were considered. One half of each set ‘‘Incorrect detection’’ represents the fact that an alarm
correspond to rocks and the other half correspond to was raised without the presence of a blocking rock. ‘‘No
no-rock images (dust, several rocks, and part of the detection’’ means that in the presence of a blocking rock
rock). In all cases, the image was scaled to a resolution the alarm was not raised.
of 20  20 pixels to be fed into the neural network.
Better results were obtained for softened images (low- Each vertical column in Figure 12 corresponds to a
pass filter). Later, a database of 1600 images was session under real conditions. Only sessions with
created, and these initial tests were considered only to significant numbers of rocks have been shown. Each
select the preprocessing technique. The extended set of column represents a working session of eight hours and
1600 images (1360 for training and 240 for testing in the the number of rocks varies depending on the day. It is
same conditions as above) were used to test results. important to note that the results are subjective, the
Table 2 shows the results for this extended set using worker supervised the system and placed the labels of
softened images. Obviously, the artificial neural network ‘‘correct-incorrect-no detection’’. It was not possible to
stop the system and measure the exact dimensions of the
Table 1. Percentage of correct identification (test set) rocks to verify the correctness of the label, but the
Normalized Equalized Low-pass
worker’s experience was taken into account.
images images filtered images
Two important observations have been noted during
FANN 93.2 86.6 93.2
in-place work: first risky situations were avoided and,
RBF 83.3 80.0 96.6
second, only one chute was stopped in a blocking
8 E. CABELLO ETAL.

Figure 12. Most significant results obtained in real conditions during nine months (from January to September, 1999) of
unsupervised work.

situation. In the traditional approach, only one worker This paper presents some results obtained using a
has to control both chutes. If one chute is blocked, he mixture of image preprocessing algorithms and neural
has to stop both chutes and crush the rocks (chutes networks. In the preprocessing phase, different algo-
could not work unsupervised, according with Spanish rithms have been tested and the final implementation
security regulations). Our system allows that one chute considered is a sequence of operators. Some of the
works without human supervision. So, from the begin- operators have been tailored to our specific problem.
ning the system helped the worker and achieved some of The output of the preprocessing phase is a set of image
the objectives. Unfortunately it is impossible to offer a windows that may contain a rock. Then, a neural
comparison in the number of stops after and before the network decides if a rock is present or not. Mechanisms
system was placed because there is no data before the avoiding false alarms due to rock grouping have been
system was placed. Initially, chutes were controlled implemented.
visually and no statistics were reported; our system is the
first automatic system placed to inspect the chutes. The Our system has been developed to face a real and
company estimates that the number of stops has been specific situation with constraints due to the system
cut by half. emplacement. But we believed that with small changes
this system could be tailored to other situations by
The results reveal a correct recognition rate of over implenting faster chutes and alternate set-up conditions.
70%, which was required by ENUSA. The system Once the system was in place, the worker felt
achieves a processing rate of 8 image/s (4 image/s per comfortable with it. Only one chute has to be stopped
chute) enough to control both chutes. when a blocking situation presents itself. On the
computer screen both chutes are displayed in real time,
and could be observed at one glance.
Conclusions
Neural network and image processing methods,
Mining industries are a very hostile environment, both working together, improve the efficiency of the system
for humans and computers. Computer systems could and the rate of false alarms is maintained at a low level.
make the work easier and avoid risky situations for Problems derived from real time, light variations and
workers. Both objectives are covered by our system. The dust in suspension have been taken into account to
computer vision system described performs the onerous achieve a robust algorithm. The system is easy to
task of estimating rock size to avoid blockages homo- operate and is working in a real environment. An
geneously and repetitively. Significant results have been extensive in-place test (nine months of unsupervised
achieved: first, risky situations are avioded, second the work) shows that objectives have been achieved.
system helps in a hard job. A computer vision system
developed for mining environments has to be robust,
and special attention has to be given to the selection of Acknowledgments
the algorithms.
The authors are grateful for the work of Julian Nieto
Different techniques have been proposed in the and Jesús M. Berrocal, both of whom were supported by
literature for dealing with rock recognition problems. Junta de Castilla y Leon. The authors would also like to
 A NEW APPROACH TO IDENTIFY BIG ROCKS WITH APPLICATIONS TO THE MINING INDUSTRY 9

thank Guido Castro and his support from Agencia 19. Looney, C.G. (1997) Pattern Recognition Using Neural
Española de Cooperación Internacional. The support of Networks. Oxford University Press.
ENUSA and its workers in Saelices (Salamanca) is also 20. Jolliffe, I.T. (1986) Principal Component Analysis. Spring-
er-Verlag.
acknowledged. 21. Fukunaga, K. (1989) Statistical Pattern Recognition. New
York: Academic Press.
22. Daugman, J.D. Complete discrete 2-D Gabor transforms
by Neural networks for image analysis and compression.
References IEEE Trans. On Acoustics, Speech and Signal Processing
36: 1169–1179.
1. Corke, P. Roberts, J. & Winstanley, G. (1998). Vision
based control for mining automation. IEEE Robotics and Appendix: Tsai’s Algorithm
Automation Magazine 5: 44–49.
2. Wang, W.X. & Stephansson, O. (1996). Automatic The number of pixels with a given gray level i is termed
selection of fragment images from a moving conveyor ni and the total number of pixels in the image is given by
belt. Mining Engineering. August: 83–88.
3. Wu, X., Devgan, A., Hagaman, R. & Kemeny, J.M. X
255
(1993). Analysis of rock fragmentation using digital image N¼ ni
processing. Journal of Geotechnical Engineering, 119: i¼0
1144–1160.
4. Crida, R.C. & De Jager, G. (1994) Rock recognition using The probability of a pixel having a given gray level is
feature classification. Proceedings of the IEEE South thus:
African Symposium on Communications and Signal Proces- ni
sing October 1994, pp. 152–157. Pi ¼
5. Crida, R.C. & De Jager, G. (1996) Multiscalar rock N
recognition using active vision. Proceedings of the IEEE The j’th moment of the gray level probability histogram
International Conference on Image Processing, September
1996.
is given by:
6. Crida, R.C. & De Jager, G. (1992). Rock detection using X
255
neural networks. Proceedings of the Third South African mj ¼ Pi ðiÞ j
Workshop on Pattern Recognition, November 1992. i¼0
7. Fernandez, R., Viennet, E., Goles, E., Barrientos, R. &
Telias, M. (1995) On-line coarse ore granulometric Note that obviously, m0=1.
analyzer using neural networks. Proceeding of ICANN
1995. For thresholding, a single threshold T will be used to
8. Russ, J.C. (1995) The Image Processing Handbook. CRC
Press. classify image pixels into two classes with mean gray
9. Baxes, G.A. (1994) Digital Image Processing. Principles level Z0 and Z1. If p0 is the fraction of pixels below T
and Applications. John Wiley. and p1 is the fraction above T, then the moments for the
10. Tsai, W.H. (1985) Moment preserving threshold: A new thresholded image are given by:
approach. Computer Vision Graphics and Image processing
29: 377–393. X
1
11. Otsu, N. (1979) A threshold selection method for gray- m0j ¼ pk ðzk Þ j
level histograms. IEEE Transactions on System, Man And k¼0
Cybernetics 9: 62–66.
12. Jain, A.K. (1989) Fundamentals of Digital Image Proces-
sing. Prentice Hall. The moments of the original image are preserved in
13. Sahoo, P.K., Soltani, S., Wong, A.K.C. & Chen, Y.C. the thresholded image by equating mj and mj (for j=0, 1,
(1988) A survey of thresholding techniques. Computer 2, 3) which produces four equalities which are known as
Vision Graphics Image Process. 41: 233–260.
14. Lee, S.U., Chung, S.Y. & Park, R.H. (1990) A compara-
the moment-preserving equations. These equations have
tive performance study of several global thresholding to be solved to obtain p0 and p1. The threshold T can
techniques for segmentation. Computer Vision Graphics then be calculated by determining the p0-tile from:
Image Process. 52: 171–190.
15. Hannah, I., Patel, D. & Davies, R. (1995) The use of X
T

variance and entropic thresholding methods for image p0 ¼ Pi

segmentation. Pattern Recognition 28: 1135–1143. i¼0
16. Vernon, D. (1991) Machine Vision. Prentice Hall.
17. Ripley, B.D. (1996) Pattern Recognition and Neural
Networks. Cambridge University Press. Since this cannot usually be evaluated exactly, the
18. Bishop, C.M. (1995) Neural Networks for Pattern Recogni- required threshold value is taken as the nearest gray
tion. Oxford University Press. level value.