Evaluation of Multilayer Perceptron and Self-Organizing Map

Neural Network Topologies applied on Microstructure

Segmentation from Metallographic Images

Victor Hugo Costa Albuquerque1, Auzuir Ripardo de Alexandria2, Paulo César Cortez3,

João Manuel R. S. Tavares1

1
Instituto de Engenharia Mecânica e Gestão Industrial (INEGI) / Faculdade de Engenharia da

Universidade do Porto (FEUP), Departamento de Engenharia Mecânica e Gestão Industrial

(DEMEGI), Rua Dr. Roberto Frias, S/N, 4200-465, Porto, Portugal

Emails: {victor.albuquerque, tavares}@fe.up.pt

2
Centro Federal de Educação Tecnológica (CEFETCE), NSMAT, Indústria, Av. Treze de

Maio, 2081, 60040-531, Fortaleza, Ceará, Brasil

Email: [email protected]

3
Universidade Federal do Ceará, Departamento de Engenharia de Teleinformática, Caixa

Postal 6007, 60.755-640, Fortaleza, Ceará, Brasil

Email: [email protected]

Corresponding author:
Prof. João Manuel R. S. Tavares
Instituto de Engenharia Mecânica e Gestão Industrial (INEGI)
Faculdade de Engenharia da Universidade do Porto (FEUP), Departamento de Engenharia
Mecânica e Gestão Industrial (DEMEGI)
Rua Dr. Roberto Frias, s/n
4200-465 PORTO
PORTUGAL
Telf.: +315 22 5081487, Fax: +315 22 5081445
Email: [email protected] Url: www.fe.up.pt/~tavares

1
Abstract

Artificial neuronal networks have been used intensively in many domains to

accomplish different computational tasks. One of these tasks is the segmentation of objects in

images, like to segment microstructures from metallographic images, and for that goal several

network topologies were proposed. This paper presents a comparative analysis between

Multilayer Perceptron and Self-Organizing Map topologies applied to segment

microstructures from metallographic images. The multilayer perceptron neural network

training was based on the backpropagation algorithm, that is a supervised training algorithm,

and the self-organizing map neural network was based on the Kohonen algorithm, being thus

an unsupervised network. Sixty samples of cast irons were considered for experimental

comparison and the results obtained by multilayer perceptron neural network were very

similar to the ones resultant by visual human inspection. However, the results obtained by

self-organizing map neural network were not so good. Indeed, multilayer perceptron neural

network always segmented efficiently the microstructures of samples in analysis, what did not

occur when self-organizing map neural network was considered. From the experiments done,

we can conclude that multilayer perceptron network is an adequate tool to be used in Material

Science fields to accomplish microstructural analysis from metallographic images in a fully

automatic and accurate manner.

Keywords: Nondestructive testing and evaluation; image processing and analysis; pattern

recognition; multilayer perceptron and self-organizing map neural networks; cast irons;

metallographic images; material sciences.

2
1 Introduction

The use of artificial neural networks is quite common in Artificial Intelligence and

Pattern Recognition domains. In particularly, they have been used in applications that involve

shapes recognition with a high degree of parallelism, considerable classification speed and

important capacity to learn through examples [1]. The smallest unit of an artificial neural

network is the artificial neuron which is a mathematical representation of a natural neuron that

compounds living neural systems.

Some domains in which artificial neural networks have been widely applied can be

found in Material Sciences. For example, they can be employed in welding control [2], to

obtain relations between process parameters and correlations in Charpy impact tests [3], to

obtain the composition of ceramic matrices [3], in the modeling of alloy elements [4], to

estimate welding parameters in pipeline welding [5], in the modeling of microstructures and

mechanical properties of steel [6], to model the deformation mechanism of titanium alloy in

hot forming [7], for the prediction of properties of austempered ductile iron [8], to predict the

carbon content and the grain size of carbon steels [9], to build models for predicting flow

stress and microstructures evolution of a hydrogenized titanium alloy [10], to perform

segmentation and quantification of microstructures in metals from images [11, 12], to classify

internal damages in steels working in creep service [13], and to perform the optimal

binarization of images in the morphological analysis of ductile cast iron [14].

Artificial neural networks have been also extremely applied to perform image

segmentation tasks. Actually, many Computational Vision systems are developed based on

artificial neural networks, essentially because of their main characteristics, like robustness to

noisily input data or outliers, execution speed and possibility to be parallel implemented.

In this work were considered two neural network topologies widely used in pattern

recognition applications: Multilayer Perceptron and Self Organizing Map, based on

3
backpropagation (supervised) and Kohonen (non-supervised) algorithms, respectively. In

particularly, these neural architectures were used and compared in the automatic segmentation

of the microstructures of metallic materials from metallographic images. To evaluate

experimentally their efficiency, it was used samples of nodular, malleable and gray cast irons.

Those materials were selected mainly due to their widespread use in industrial applications, as

in, for example, machine base structures, lamination cylinders, main bodies of valves and

pumps and gear elements.

This paper is organized as follows. In next section, some essential aspects of

Multilayer Perceptron neural networks are reviewed and the topology used is presented. The

equivalent is done in third section considering Self Organizing Map neural networks. In next

section the experimental work accomplished is explained and results are discussed. In last

section the main conclusions are presented.

2 Multilayer Perceptron Neural Network

A human brain is composed by about ten billion neurons and their organization is of

high structural and functional complexity. These units are densely interconnected, which

results in a very complex architecture and with an intelligence level that was not yet achieved

by any artificial system. Several mathematical models have been developed to represent

neurons and their interconnection. In this direction, artificial neural networks have appeared

as an attempt to reproduce human brain potentialities, specially its learning ability [15].

The first neuron mathematical model was proposed by McCulloch and Pitts [16]. It

has a binary output and several inputs, each one with a different excitatory or inhibitory gain.

These gains are known as synaptic weights (or only weights). Input signal values and

associated weights determine the neuron output [15].

4
 Thus, perceptron or artificial neuron is the mathematical model of a neuron cell and

the basic unit that compounds an artificial neural network. Perceptron architecture consists of

a set of n inputs (xi), each one associated to a weight (wi) and an activation function (fi).

Perceptrons can be organized to form a layer, in which all perceptrons are linked to the same

inputs but have distinct outputs. This kind of network is designated as a perceptron network.

Perceptron networks can achieve good performances only when the pattern to be recognized

is linearly separable [15]; therefore, they should not be used to solve complex classification

problems involving non-linearly separable patterns, instead multilayer perceptron networks

can be used.

2.1 Multilayer perceptron networks architecture

Multilayer perceptron networks are formed by an input layer (Xi), one or more

intermediary or hidden layers (HL) and an output layer (Y). A weight matrix (W) can be

defined for each of these layers. This artificial neuronal network topology can solve

classification problems involving non-linearly separable patterns and can be used as a

universal function generator [15].

Multilayer perceptron networks have two distinct phases: training and execution. With

this network topology, it is impossible to use directly the usual delta rule [15] for the training

phase, since this rule does not permit weight recalculation for hidden layers. Therefore, the

most widely used algorithm for multilayer perceptron networks training is the

backpropagation and its variants. This learning approach is more complex than the one for a

perceptron network and it is of the type supervised [16].

2.1.1 Backpropagation Algorithm

5
 Standard backpropagation algorithm is one of the most used algorithms to train

neuronal networks [15, 17]. This algorithm is applied by initializing the weights randomly

and presenting examples to the neuronal network as an input signal. Then, this signal is

propagated through the hidden layers and achieves the output layer where the network outputs

are obtained. Afterwards, the weights (w) are modified according to equation:

∆wij = η . δ j (k ) sj (k ) , (1)

d j (k ) − s j (k ), output layer and

where δ j (k ) = f i .elocalj to elocalj =  ,
 ∑ pw jp .δ p ,
l −1
otherwise

and i is the current neuron input, j is the neuron, l is the current layer and, s is neuron output,

d is the desired output, fi is a derivative activation function. Finally, this procedure is repeated

until the error is less than a previously adopted threshold value.

This algorithm presents some disadvantages like, for example, during the search for a global

minimum, the surface where the error derivative is almost null can be achieved and remain

there forever. The convergence of the neural network can be very time-consuming since the

trajectory of the initial weights set towards the final one is randomly defined.

2.2 Topology of the multilayer perceptron network used

In this work was used a multilayer perceptron with two layers. Since the main task was

the segmentation of material microstructures from metallographic images, it was used to

classify the pixels of a metallographic image as belonging to one of these material

microstructures: graphite, pearlite or ferrite. For that, we choose a 3/2/3 (three inputs, two

perceptrons in hidden layer and three perceptrons in output layer) topology. For each input

was assigned one color component (R, G and B) of each pixel of the input image to be

segmented. Each output was assigned to the three possible pixel classifications: graphite,

pearlite or ferrite. The number of perceptrons used in the hidden layer was established using

6
the method proposed by Yin, Liu and Han [18]. Finally, to perform microstructures

quantification, for each microstructure segmented, the associated pixels were counted and the

relative percent value for the whole image was calculated.

As already referred, to train the artificial network was used backpropagation algorithm

and the examples considered were microstructure pixels chosen from representative input

images. It should be noted that the training phase just needed to be done once for each kind of

metallic material to be segmented.

3 Self-organizing map neural networks

Self-organizing map neural network is a topology proposed by Kohonen [19] that

consists in a feedforward neural network that uses a non-supervised training. This model is

usually built by a neuron layer disposed linearly (1D) or through a plane (2D) and is classified

as a self-organizing map since the neurons are disposed in an one-dimensional or bi-

dimensional reticulated, as shown in figure 1. In this figure is illustrated a self-organizing

neural network with a 1D topology and a bi-dimensional network in a squared topology with

nine neurons organized in a 3x3 array.

When a pattern is presented to a self-organizing map neural network, their neurons

compete among themselves in order to determine the one that has the best response to the

input pattern. The neuron that generates the least Euclidian distance between the input and

weights vector is the chosen winner, and it has its weights adjusted to respond better to the

input pattern. Kohonen network, then, has the characteristic of modifying internally so that

the neurons respond better to a determined input pattern [20].

Concerning self-organizing non-supervised learning model, not only the winner

neuron but also the ones in its neighborhood have their weights adjusted, as is explained in

the next section. This paradigm is based on the theory that the cells from brain cortex are

7
anatomically arranged in function of its stimulus originated from sensors that are directly

connected to them [20].

Each neuron of a self-organizing neural network represents an output. Another

characteristic of this network topology is that its neurons are totally connected (synaptic

links), so that if there are ten inputs, each neuron has ten inputs, each one linked to a point in

the input layer. Then, each synaptic connection has a weight so that if they are ten input

points, and should be three neurons in the network, then we should be thirty connections (ten

for each neuron) and consequently, thirty synaptic weights (ten for each neuron).

3.1 Neighborhood

A neuron in a self-organizing map network has a set of neighbors that can be

topologically organized in regions to have the best response to a given stimulus. This

characteristic is similar to the one that occurs in human brain, where, depending on the

activity that is happening, there are some neural centers of more intensified activity.

In the beginning of the training step, the set of neighbors is large, but it decreases

along time, while the network organizes itself. In fact, for its better organization, the

neighborhood begins wide and decreases monotonically, because if the neighborhood begins

very small, comparing to the whole map, the network does not become globally in order [20].

Usually, at the end of the process, neighborhood radius must be null and only the winner has

its weights adjusted.

Neighborhood weight adjustment permits that neurons near to the winner have good

conditions to dispute with the same during next iterations, improving in this way the

competition for the best network learning.

3.2 Neuron adaptation

8
 Neuron adaptation consists in weight adjustment, looking for response improvement to

determined stimulus. This process is fundamental for the formation of an ordered network.

The first step to achieve weight adjustment is to determine the winner neuron. So, the winner

will be the neuron that presents the least Euclidean distance dt(n) between an input pattern and

its weights, calculated as:

I
dt ( n ) = ∑ ( xt ( i ) − wn ,t ( i ))2 , (2)
i =1

where xt(i) is the ith input vector component at time t, wn,t(i) is the ith weights vector

component of a neuron n at time t, i is the input and weight index of a network with I inputs

and dt(n) is Euclidean distance of neuron n at time t.

Euclidian distance, as presented in equation 2, is the sum of squared differences

between each input and its correspondent weight. In fact, the least Euclidean distance

represents the neuron which weight value is the most similar to input value presented [20]. In

this way, in each iteration, the neuron that has the least Euclidian distance and its neighbors

must have their weights adjusted to have a better response to that input, while the other

neurons remain equal. Weight adaptation at next instant t + 1 is a simple process. It consists in

taking the difference between vectors X (input) and W (weights) and sum a fraction of this

difference to the weight vector at time t (actual). Neighbor weights are adjusted using the

same principle, expressed by:

wn (t + 1) = wn (t ) + as (t ).( x(t ) − wn (t )) 2 , (3)

where as(t) is the learning rate and corresponds to a fraction of difference between X and W

that is summed to W. So that, 0 < as < 1 [15].

Usually as(t) (equation 3) is initialized with a high value (around one), and

successively decreases, following any rule or function, until its value achieves approximately

9
zero. Then, map ordering that takes place in the beginning of training is affected by a large

neighborhood radius and by a high as(t) value.

3.3 Iterations

The number of iterations or epochs corresponds to the number of times that the input

patterns are presented to the network. On each epoch, it must be presented all input data. For

each pattern a winner neuron is determined. Then, its weights and the weights of its neighbors

are adjusted. After the last pattern is presented, a new epoch begins and all input data must be

presented again [19].

3.4 Training algorithm for a self-organizing map neural network

The self-organizing training algorithm used can be described as follows [19]:

1. specify the number of epochs;

2. initialize the network weights using random values;

3. specify the initial neighborhood radius for each neuron (it is recommended to use

the whole network as initial radius);

4. present an input to the network;

5. calculate the output for each neuron, using Euclidean distance between input and

network weight;

6. select the winner neuron;

7. adjust winner weights and the ones of its neighbors;

8. if still there are neighbors, then decrease neighborhood radius;

9. if still there are input data, then go to step 4;

10. increment number of epochs and if allowed maximum number of epochs was still

not achieved, then go to step 4 and present all input data again.

10
 It is important that the initial weight values are different from each other and much

smaller than input data. If the weights are initialized with values similar to input data, then the

network has the tendency to choose specific winners, avoiding its self organization.

3.5 Winner neuron

After network training, we can label each neuron according to the pattern it classifies.

By this way, if, for example, a network is classifying objects, and the winner neuron

represents a determined object, this neuron can be labeled with that object name.

Nevertheless, many neurons can classify the same object. In fact, this can occur because of its

neighborhood situation. If this case happens, then there will be more than a neuron with the

same label and a post-processing step can be necessary.

3.6 Topology of the self-organizing map neural network used

The applied self-organizing map topology was a 1-D self-organizing network and its

main task was to perform the segmentation of material microstructures from metallographic

images. Thus, self-organizing map neuronal network was used to recognize the pixels of a

metallographic image that belong to each of these material microstructures: graphite, pearlite

or ferrite. Hence, we choose a three neurons network having three inputs on each of them. To

each input is assigned one color component of input image pixels (R, G and B). Each neuron

is assigned to three possible pixel classifications: graphite, pearlite or ferrite. To perform

microstructures quantification, segmented pixels were counted and the associated percent was

calculated. Kohonen training algorithm was used to train the network and the examples used

were microstructure pixels obtained automatically from representative input images.

11
4 Results and discussion

The aim of this work was to segment material microstructures from metallographic

images; for this, we used two types of artificial neural network topologies to identify the

following microstructures: graphite, pearlite and ferrite. The experimental results were

obtained applying multilayer perceptron and self-organizing map neural networks on samples

of nodular, malleable and gray cast irons. Here we present a comparative analysis between

these two artificial neural network topologies in the automatic segmentation of metal

microstructures from metallographic images. For that comparison, we consider the analysis

presented by the Albuquerque et al. [12], which validated the quality of the segmentation and

quantification results obtained by the multilayer perceptron network on the microstructures of

cast irons from metallographic images. In the referred work, the results obtained by the

multilayer perceptron network were very similar to the ones obtained by visual human

inspection and better than the results obtained using a commercial system that is awfully

common and well accepted in the microstructures analysis area. Thus, we adopted the results

obtained by the multilayer perceptron network as a reference basis.

In the previous sections, we describe the neuronal network topologies to identify three

different microstructures in input images, however both networks were only used to identify

two classes, where one class indicates the presence of graphite (associated with black color in

the input images) and the other class refers to the presence of pearlite or ferrite (associated

with the gray color in the input images). The reason for the association pearlite/ferrite in one

class was because we were not able to visually distinguish these two microstructures in the

input images, which could affect the further results analysis.

The experimental images considered here had size of 640 x 480 pixels, the training

phase was performed once by an experimented operator considering four representative

images of each iron considered. This number of training images was defined in function of the

12
number of microstructures involved, and the training images were manually elected by the

operator through visual inspection. The same operator accomplished the experimental results

analysis and comparison. It should be noticed that different results could be obtained if the

networks were trained using different sets of training images. In other words, the quality of

the results accomplished depends on the quality of the image training set used and of each

image considered. That is, if all microstructures involved are or not properly represented in

the training images, the samples were or not well metallographic prepared or the images were

or not appropriately acquired as, for example, using inadequate contrast and illumination

conditions, the results quality can be affected.

In the training of the multilayer perceptron network, the backpropagation algorithm

was employed, adopting a learning rate of 0.1, a moment rate of 0.001, and as stopping

criterion the number of epochs equal to 2500 or an absolute error not greater than 0.01. In the

training of the self-organizing map network was used the Kohonen algorithm, adopting a

learning rate of 0.1 and as stopping criterion the number of epochs equal to 2500.

In table 1, are shown the results obtained using multilayer perceptron and self-

organizing map neural networks in the case of the nodular cast iron. It should be noticed that

the results obtained are similar, presenting a minimum difference of 1.00% and a maximum of

5.59%, for samples 20 and 17, respectively. Moreover, it can be noticed that the multilayer

perceptron network presented an average of graphite equal to 12.09% and pearlite or ferrite to

87.91% and that self-organizing map network presented 15.35% and 84.65% of graphite and

pearlite or ferrite, respectively, being the difference average equal to 3.26%.

Figures 2 (a), (b) and (c), present an original image of a nodular cast iron, and the

images resulting of the segmentation process accomplished using multilayer perceptron and

self-organizing map neuronal networks, respectively. These images are the ones that

presented the major difference in the results obtained by the two neuronal networks.

13
Additionally, images of figures 2 (d), (e) and (f), are the ones that presented the minor

difference.

It should be noticed that the segmentation done using multilayer perceptron network is

more accurate than the one obtained using the self-organizing map network, since part of the

pearlite or ferrite was erroneously segmented and classified as graphite by self-organizing

map network. This fact is the main justification for the existing difference in the results

obtained.

In table 2, the results of the segmentation done using multilayer perceptron and self-

organizing map networks on samples of a gray cast iron are presented. These results were the

most dissimilar for the studied irons, presenting a minimum difference of 1.86% and

maximum of 14.70%, for samples 15 and 18, respectively. Moreover, it can be noticed that

multilayer perceptron network presented an average of graphite equal to 10.28% and of

pearlite or ferrite to 89.62%, and that self-organizing map network presented 19.62% and

80.38% of graphite and of pearlite or ferrite, respectively, being the difference average

between the two neural networks equal to 9.24%.

Figures 3 (a), (b) and (c), present the original image of a gray cast iron, and the

resulting images of the segmentation done using multilayer perceptron and self-organizing

map neural networks, respectively. These images are the ones that presented the minor

difference between results obtained by the two neural networks. Additionally, images of

figures 3 (d), (e) and (f), are the ones that presented the major difference.

Analyzing the results obtained, one can conclude that the segmentations accomplished

using multilayer perceptron and self-organizing map neural network topologies are similar for

sample 1. Relatively to sample 11, the segmentation done by visual inspection is very similar

to the one obtained using multilayer perceptron network which not occurs with self-

organizing map network, because pearlite or ferrite was erroneously segmented as graphite.

14
 In table 3, the results obtained by the segmentation done using multilayer perceptron

and self-organizing map methods on a malleable cast iron are presented. These are the most

similar results obtained by the two neural networks on all samples analyzed, presenting a

minimal difference of 1.54% and maximum of 5.62% for samples 3 and 19, respectively.

Moreover, it can be noticed that multilayer perceptron network presented an average of

graphite equal to 14.98% and of pearlite or ferrite to 85.02% and self-organizing map network

presented 17.70% and 82.30% of graphite and pearlite or ferrite, respectively, being the

involved difference average between the neural networks equal to 2.72%.

Figures 4 (a), (b) and (c), present an original image of a malleable cast iron and the

images resulting from the segmentation done using multilayer perceptron and self-organizing

map neural network, respectively. Notice that these images presented the minor difference

verified between the results obtained by the used networks. Additionally, the images of

figures 4 (d), (e) and (f), presented the major difference.

Verifying the results obtained, it can be noticed that the segmentations performed

using multilayer perceptron and self-organizing map neural networks are analogous on

sample 3. For sample 19, the segmentations obtained are distinct, because self-organizing

map network segments great part of pearlite or ferrite as graphite (figure 4(f)). However, that

error is not verified when used multilayer perceptron network that segmented correctly the

graphite from the other two constituents.

In the results obtained by self-organizing map network on samples of nodular, gray

and, in particularly, malleable casting iron, we can verify its considerable difficulty to

segment graphite successfully when the background of the input image was not uniform.

However, multilayer perceptron network did not show this difficulty, and so it could get

segmentation results very effectively.

15
 Multilayer perceptron neural network showed to be a versatile and easy to use solution

to perform automatic segmentation of material microstructures from metallographic images,

even when the input images were of low quality. Moreover, when compared with self-

organizing map network, solution network needed less time to accomplish the segmentation

of the structures presented. Self-organizing network had as main advantage its training

algorithm that was easier and faster.

5 Conclusions

This paper described two neural network topologies here considered to perform the

segmentation of metallic material constituents from images. The neural network solutions

were based on multilayer perceptron and self-organizing map neural topologies and use

backpropagation and Kohonen algorithms. These solutions are more robust to noisy input and

illumination irregularities during image acquisition than traditional segmentation techniques,

as usual threshold approach, for example.

In this work was accomplished a comparative analysis on the experimental results

obtained using two network topologies in the segmentation of microstructures from

metallographic images of nodular, malleable and gray cast irons. The one that showed better

results was the multilayer perceptron network.

From the experimental results accomplished, we can conclude that multilayer

perceptron network can be successfully used in applications of Materials Science fields; in

particular, for the segmentation of material microstructures from metallographic images.

Comparatively to self-organizing map network, multilayer perceptron network presents as

main advantages the reduction of the segmentation time and results of higher quality.

References

16
[1] S. Samarasinghe, Neural networks for applied sciences and engineering: from

fundamentals to complex pattern recognition, Auerbach Publications, 2006.

[2] H.K.D.H. Bhadeshia, Neural networks in materials science, ISIJ International 39(10): 966-

979, 1999.

[3] H.K.D.H. Bhadeshia, Neural networks and genetic algorithms in materials science and

engineering, Tata McGraw-Hill Publishing Company Ltd., India: 1-9, 2006.

[4] L. Miaoquana, X. Aiming, H. Weichao, W. Hairong, S. Shaobo and S. Lichuang,

Microstructural evolution and modelling of the hot compression of a TC6 titanium alloy,

Materials Characterization 49(3): 203-209, 2004.

[5] I. Kim, Y. Jeong, C. Lee and P. Yarlagadda, Prediction of welding parameters for pipeline

welding using an intelligent system, The International Journal of Advanced Manufacturing

Technology 22(9-10): 713-719, 2003.

[6] J. Kusiak and R. Kusiak, Modelling of microstructure and mechanical properties of steel

using the artificial neural network, Journal of Materials Processing Technology 127(1): 115-

121, 2002.

[7] X. Xiao-li and L. Miao-quan, Microstructure evolution model based on deformation

mechanism of titanium alloy in hot forming, Transactions of nonferrous metals society of

China 15(4): 749-753, 2005.

[8] R. Biernacki, J. Kozłowski, D. Myszka and M. Perzyk, Prediction of properties of

austempered ductile iron assisted by artificial neural network, Materials Science 12(1): 11-15,

2006.

[9] A. Abdelhay, Application of artificial neural networks to predict the carbon content and

the grain size for carbon steels, Egyptian Journal of Solids 25(2): 229-243, 2002.

17
[10] O. Wang, J. Lai and D. Sun, Artificial neural network models for predicting flow stress

and microstructure evolution of a hydrogenized titanium alloy, Key Engineering Materials

353-358: 541-544, 2007.

[11] V.H.C. Albuquerque, P.C. Cortez, A.R. Alexandria, W.M. Aguiar and E.M. Silva, Image

segmentation system for quantification of microstructures in metals using artificial neural

networks, Revista Matéria 12(2): 394-407, 2007.

[12] V.H.C. Albuquerque, P.C. Cortez, A.R. Alexandria and J.M.RS. Tavares, A new solution

for automatic microstructures analysis from images based on a backpropagation artificial

neural network, Nondestructive Testing and Evaluation 23(4): 273-283, 2008.

[13] L.A. Dobrzanski, M. Sroka and J. Dobrzanski, Application of neural networks to

classification of internal damages in steels working in creep service, Journal of Achievements

in Materials and Manufacturing Engineering 20(1-2): 303-305, 2007.

[14] A. de Santis, O. di Bartolomeo, D. Iacoviello, F. Iacoviello, Optimal binarization of

images by neural networks for morphological analysis of ductile cast iron, Pattern Analysis

and Applications 10(2): 125-133, 2007.

[15] S. Haykin, Neural Networks and Learning Machines. Prentice Hall, USA, 2009.

[16] W.S. McCullogh and W. Pitts, A logical calculus of the ideas immanent in nervous

activity, Neurocomputing: foundations of research book contents: 15-27, 1988.

[17] D. Plaut, S.J. Nowlan and G.E. Hinton, Experiments on learning by backpropagation.

Computer Science Department, Carnegie – Mellon University, Technical Report CMU-CS,

1986.

[18] X.C. Yin, C.P. Liu and Z. Han, Feature combination using boosting. Pattern Recognition

Letters 26(16): 2195-2205, 2005.

[19] T. Kohonen, Self-organized formation of topologically correct feature maps, Biological

Cybernetics 43(1): 59-69, 1982.

18
[20] T. Kohonen, Self-Organizing Maps. Springer Series in Information Sciences 30,

Springer-Verlag, Berlin, 2001.

19
FIGURES CAPTION

Figure 1: Two models for self-organizing map networks: a) 1-D and b) 2-D.

Figure 2: Two original images of a nodular cast iron, a) and d); resultant segmentation using

multilayer perceptron, b) and e), and self-organizing map, c) and f), neural networks.

Figure 3: Original images of a gray cast iron, a) and d); resultant images of the segmentation

done using multilayer perceptron, b) and e), and self-organizing map, c) and f), neural

networks.

Figure 4: Original image of a malleable cast iron, a) and d); resultant images of the

segmentation done using multilayer perceptron, b) and e), and self-organizing map, c) and f),

neural networks.

20
TABLES CAPTION

Table 1: Results obtained using multilayer perceptron (MLP) and Kohonen self-organizing

map based neural networks on samples of a nodular cast iron.

Table 2: Results obtained using multilayer perceptron (MLP) and Kohonen self-organizing

map (SOM) based neural networks on samples of a gray cast iron.

Table 3: Results obtained using multilayer perceptron (MLP) and Kohonen self-organizing

map (SOM) based neural networks on samples of a malleable cast iron.

21
Acknowledgments

To Federal Center of Technological Education of Ceará - CEFET CE, for the support

given for the accomplishment of this work, in particular to Mechanical Testing Laboratory

and to Teleinformatic Laboratory. The authors would like to thank also to CAPES for their

financial support.

22
FIGURES

Figure 1a

Figure 1b

Figure 2a

23
Figure 2b

Figure 2c

Figure 2d

24
Figure 2e

Figure 2f

Figure 3a

25
Figure 3b

Figure 3c

Figure 3d

26
Figure 3e

Figure 3f

Figure 4a

27
Figure 4b

Figure 4c

Figure 4d

28
Figure 4e

Figure 4f

29
TABLES

Table 1

Nodular cast iron

Samples MLP network (%) SOM network (%)

Graphite Ferrite/Pearlite Graphite Ferrite/Pearlite
1 11.51 88.49 14.58 85.42
2 13.36 86.64 16.28 83.72
3 13.19 86.81 15.80 84.20
4 13.46 86.54 16.34 83.66
5 12.46 87.54 14.80 85.20
6 11.79 88.21 15.47 84.53
7 14.58 85.42 16.88 83.12
8 12.50 87.50 14.37 85.63
9 14.02 85.98 15.80 84.20
10 13.30 86.70 15.55 84.45
11 9.08 90.92 14.04 85.96
12 9.25 90.75 14.07 85.93
13 11.56 88.44 16.33 83.67
14 9.24 90.76 13.80 86.20
15 10.22 89.78 15.26 84.74
16 9.89 90.11 14.52 85.48
17 8.31 91.69 13.90 86.10
18 7.66 92.34 13.21 86.79
19 14.66 85.34 13.28 86.72
20 21.73 78.27 22.73 77.27
Average 12.09 87.91 15.35 84.65

30
 Table 2

Gray cast iron

Samples MLP network (%) SOM network (%)

Graphite Ferrite/Pearlite Graphite Ferrite/Pearlite
1 12.21 87.79 16.37 83.63
2 6.80 93.20 12.70 87.30
3 8.56 91.44 16.77 83.23
4 6.80 93.20 13.79 86.21
5 8.03 91.97 18.44 81.56
6 5.02 94.98 19.16 80.84
7 8.73 91.27 17.85 82.15
8 9.15 90.85 13.87 86.13
9 7.90 92.10 14.41 85.59
10 6.67 93.33 14.40 85.60
11 40.41 59.59 49.27 50.73
12 7.47 92.53 19.38 80.62
13 9.83 90.17 17.36 82.64
14 7.52 92.48 22.72 77.28
15 21.69 78.31 23.55 76.45
16 8.79 91.21 19.00 81.00
17 7.76 92.24 21.29 78.71
18 6.77 93.23 21.47 78.53
19 7.64 90.28 19.73 80.27
20 7.78 92.22 20.86 79.14
Average 10.28 89.62 19.62 80.38

31
 Table 3

Malleable cast iron

Samples MLP network (%) SOM network (%)
Graphite Ferrite/Pearlite Graphite Ferrite/Pearlite
1 18.95 81.05 20.57 79.43
2 15.71 84.29 17.73 82.27
3 14.96 85.04 16.50 83.50
4 14.00 86.00 15.63 84.37
5 15.16 84.84 17.28 82.72
6 16.07 83.93 18.32 81.68
7 19.11 80.89 20.76 79.24
8 19.22 80.78 22.17 77.83
9 15.64 84.36 17.60 82.40
10 17.64 82.36 19.40 80.60
11 11.84 88.16 16.82 83.18
12 12.04 87.96 17.32 82.68
13 13.72 86.28 18.84 81.16
14 14.06 85.94 18.07 81.93
15 11.83 88.17 16.07 83.93
16 11.40 88.60 16.19 83.81
17 11.30 88.70 15.54 84.46
18 10.68 89.32 15.15 84.85
19 21.05 78.95 15.43 84.57
20 15.27 84.73 18.57 81.43
Average 14.98 85.02 17.70 82.30

32