A New Tool for Automatic Classification of Microstructure Based on Backpropagation Artificial Neural Network

This paper presents a comparative analysis between two automatic systems to segment and quantify microstructures of nodular cast iron, malleable cast iron, and gray cast iron, in images. The compared computational systems are the SVRNA system (Computational Vision System based on an Artificial Neural Network for Microstructure Segmentation and Quantification), developed during this work and here presented, that uses an artificial neural network based on the backpropagation algorithm to segment and quantify microstructures in metallic materials, and the Image Pro-Plus, a common tool used for the same purpose. For results comparison 60 samples of cast iron had been considered and analyzed, and the results of our SVRNA system are very similar to the ones obtained by visual human inspection. In fact, the SVRNA system has segmented efficiently and automatically the microstructures of the materials in analysis, what has not always occurred with the Image Pro-Plus. Thus, we could conclude that our SVRNA system is a valid and adequate option for researchers, engineers, specialists and others of Material Sciences field accomplish microstructural analysis from images in a fully automatic and efficient manner

For Peer Review Only

SVRNA System A New Tool for Automatic Classification of

Microstructure Based on Backpropagation Artificial Neural Network

Journal: Nondestructive Testing and Evaluation

Manuscript ID: GNTE-2007-0002

Manuscript Type: Original Article

Date Submitted by the

Author: 15-Dec-2007 Complete List of Authors: Albuquerque, Victor; Universidade do Porto

Cortez, Paulo; Universidade Federal do Ceará, Departamento de Engenharia de Teleinformática

Alexandria, Auzuir; Universidade Federal do Ceará, Departamento

de Engenharia de Teleinformática Tavares, João; Universidade do Porto Keywords: artificial neural networks, image processing and analysis, image

segmentation and quantification, Micrstructure classification

For Peer Review Only

SVRNA system – a new tool for automatic classification of microstructure based

on backpropagation artificial neural network

V H C de Albuquerque 1*, P C Cortez2, A R de Alexandria2, João Manuel R S

Tavares1

1Instituto de Engenharia Mecânica e Gestão Industrial (INEGI), Faculdade de Engenharia da Universidade do Porto (FEUP), Departamento de Mecânica e Gestão Industrial (DEMEGI), Rua Dr Roberto Frias, S/N, 4200 - 465 Porto – Portugal

2Universidade Federal do Ceará (UFC), Departamento de Engenharia de Teleinformática, Avenida Mister Hull, S/N – Pici, CEP 60455.970 - Fortaleza – Ceará,

For Peer Review Only

Abstract

This paper presents a comparative analysis between two automatic systems to segment and quantify microstructures of nodular cast iron, malleable cast iron, and gray cast iron, in images The compared computational systems are the SVRNA system (Computational Vision System based on an Artificial Neural Network for Microstructure Segmentation and Quantification), developed during this work and here presented, that uses an artificial neural network based on the backpropagation algorithm to segment and quantify microstructures in metallic materials, and the Image Pro-Plus, a common tool used for the same purpose For results comparison 60 samples of cast iron had been considered and analyzed, and the results of our SVRNA system are very similar to the ones obtained by visual human inspection In fact, the SVRNA system has segmented efficiently and automatically the microstructures of the materials in analysis, what has not always occurred with the Image Pro-Plus Thus, we could conclude that our SVRNA system is a valid and adequate option for researchers, engineers, specialists and others of Material Sciences field accomplish microstructural analysis from images in a fully automatic and efficient manner

Key-words: Artificial neural networks; Image processing and analysis; Image

segmentation and quantification; Microstructure; Multilayer Perceptron; Materials Science

*Corresponding author Email: dce06017@fe.up.pt

For Peer Review Only

1 Introduction

One of the greatest existing challenges in the development of machines and equipments is the conception of systems with intelligent capacities, so as they can consider problems in an analog way to humans Thus, various advances have occur in the domain of Artificial Intelligence, which has as main objective the research of new algorithms and technological solutions for the construction of systems with intelligent capacities However, this is a very difficult and complex task, as there are innumerable tasks that humans commonly and naturally do as, for example, visualize, hear, walking and speech, but that are not trivial process to implement in computational systems

Among the tasks above cited, the vision has deserved special attention of the scientific community; in particular, because of the considerable number of existing applications and its major importance in our society [1] Initially derived from Artificial Intelligence field, Computational Vision became a distinct research area that searches to build computational systems able to perform some visual information analysis and interpretation, which can be capable to help humans to execute some usual tasks with higher speed and precision [2] This new area uses, between others, techniques from Artificial Intelligence, Digital Signal Processing and Pattern Recognition fields [3]

Artificial Neural Networks (ANN) is one of the techniques used in Artificial Intelligence and Pattern Recognition fields In particularly, they have been used in applications that involve Pattern Recognition with high degree of parallelism, considerable classification speed and important capacity to learn through examples [4]

For Peer Review Only

Artificial Neural Networks have been used in some domains of Materials Science as well Some examples that can be found are: in welding control [5], to obtain the relations between process parameters and correlations in Charpy impact tests [6], to get the composition of models for ceramic matrices [6] , in the modeling of alloy elements [7,8], prediction of welding parameters in pipeline welding [9], modelling of microstructure and mechanical properties of steel [10], model based on deformation mechanism of titanium alloy in hot forming [11], prediction of properties of austempered ductile iron [12], predict the carbon content and the grain size of carbon steels [13], models for predicting flow stress and microstructure evolution of a hydrogenized titanium alloy [14], among others

In this work, the SVRNA system developed in the scope of this work is present

This system is able to accomplish the quantification of microstructures in metallic materials from images Also, in this work, the SVRNA system is compared with the Image Pro-Plus system, which is a image analysis software for fluorescence imaging, materials imaging, and various other scientific areas as, for example, biomedical engineer and industrial applications [15] For this comparison, we used experimental samples of nodular, malleable and gray cast iron These materials were selected manly because of their large use in industry, as in, for example, machine base structures, lamination cylinders, main bodies of valves and pumps, gear elements, among others

2 Materials and methods

In this work, the materials considered to be analyzed are nodular cast iron, malleable cast iron and gray cast iron, which present microstructures with different shapes

For Peer Review Only

Thereafter, we need the application of computational techniques to automatically segment and quantify the constituents of the cast irons considered that are presented in images acquired for that purpose To accomplish this, it has developed the already referred SVRNA system, which is based on an Artificial Neural Network (ANN)

As the nervous system, ANN do not possess discrete functional components However, for a determined task, there are responsible regions with high degree of redundancy and parallelism, which turn the computational systems based on ANN usually robust and fast

The ANN are composed by a set of perceptrons, being the perceptron a model of a nervous cell (neuron) Such element is the most basic one that we can find on an ANN [16] Perceptron networks are neural networks of a unique layer that recognize only linear separable patterns; in general, these networks are only used in simple tasks

For the solution of more complex problems, the networks MLP (MultiLayers Perceptron) are used, whose pattern, in general, are not linear separable Usually, these networks consist of an input layer, several hidden layers and an output layer The same nets possess two distinct phases: training and execution In this kind of neural networks is widely used the backpropagation algorithm, as well as are its variants [17] The learning rule of this kind of networks is considerably more complex than the rule used for simple perceptron networks This learning is also known as "supervised", because it is necessary to know previously the correct pattern so that the training is carried through successfully, needing an adequate operator for that [17]

For Peer Review Only

2.1 Cast Iron

Cast irons are composed by iron and carbon and possess contents of this last one between 2.14 and 6.70% In fact, the majority of cast irons contain between 3.0 and 4.5% of carbon [18] Beside carbon, the cast irons present significant silicon contents, and, therefore, some authors as, for example, Callister [18], Chiaverini [19] and Raabe et al

[20], consider the cast iron as an iron alloy, with carbon and silicon Another characteristic

of the cast irons is associated with the form that carbon presents in its microstructure For example, in gray cast irons, all the carbon is presented practically free in the form of lamellar graphite or in veins, while that in the white cast iron, the carbon presented is arranged in the form of cementite (Fe3C) [21]

In accordance with the carbon content, cast irons can be classified in hipoeutectics, eutectics or hipereutectics, as it is shown in Figure 1 When eutectic cast iron is solidified, just below of point G (see Figure 1), it happens the formation of a microstructure with cementite and austenite globules, called ledeburite I Carrying on the cooling process, below the 727 ºC line, the austenite transforms itself into pearlite globules on the cementite, forming ledeburite II [22], that was particularly considered in this work

Line I of the iron-carbon diagram shown in Figure 1, specifies the position of a hipoeutectic cast iron or, using other words, with carbon content inferior than 4.3% Line II,

in turn, specifies a hipereutetic cast iron, that presents carbon content superior than 4.3%

For Peer Review Only

3 SVRNA system

Our SVRNA system was implemented in C++ using the environment Builder 6 from Borland, runs in Microsoft Windows XP platforms and, as already referred, uses an artificial neural network, trained using the backpropagation algorithm, to accomplish the segmentation and quantification of microstructures in metallographic images More specifically, an artificial neural network composed by 42 neurons distributed in two layers

is used The neuron distribution along the ANN layers (topology) corresponds to: 3 inputs,

30 neurons in the hidden layer and 9 neurons in the output layer The inputs for the neural network are the R (red), G (green) and B (blue) components, or the gray level instead, of pixels selected in the input image The output of the same network is the indication of the region, which must be associated to each pixel considered, through the indication of the pseudocolor that must be attributed in its labeling

As previously described, the neural network integrated in our SVRNA system is used to segment the original images acquired for the analysis of microstructures, being their classification and quantification made from the pseudocolors associated to the same ones The training of the network used is made considering a reduce sample set of pixels from each microstructure to consider, attributing to each set a pseudocolor that is used as the label of the associated microstructure Thus, on the training of the neural network employed are used example sets of components R, G, B, or of correspondent gray levels, and the associated pseudocolors (numbered from 0 to 8)

For Peer Review Only

The developed system interacts with its user in order to collect the pixels of the selected example sets, considering for that the computer mouse These example sets are then used in the training phase of the neural network and to obtain the correspondent microstructures classification With an adequate training of the network used, the same network can be used in the classification and quantification of similar images of the identical materials

The following computational modules constitute our SVRNA system: training, manipulation of records, execution and output The training module is responsible, as its proper name indicates, for the learning phase of the neural network used, being the backpropagation algorithm considered for that of supervised type The functions of the manipulation of records module are: the reading and generation of new records The execution module of SVRNA is responsible for the segmentation of the microstructures present in the original input image and, consequently, for its classification This module just can be used after the training process of the neural network used, and it is initiated with the opening of the record associated to an input image and it is finished with the display of the segmented image and the obtained numerical results, using for that an area chart with the percentage of each classified microstructure The segmented images and the area chart are the outputs of the segmentation process carried through the neural network used, and can be visualized and saved for posterior consideration

For Peer Review Only

To segment and quantify the microstructures present in material samples using our SVRNA system, is necessary to carry through a process of samples preparation: metallography In other words, we need to section, countersink, sand, polish and etch each sample to the be analyzed After that, the images of each material sample are acquired using a digital camera direct-coupled to an optical microscope

In Figure 2, is presented the interface of SVRNA system In this image, we can visualize an original image of a gray cast iron and the corresponding segmented image obtained using our SVRNA system

The use of our SVRNA system is very practical and intuitive, and it consists in the following First, the user opens the original image to be analyzed Then, the user resets the neural network and afterward, for each microstructure to be segmented, selects the desired pseudocolors and initiates the selection process of some pixels that should be considered by the network for the posterior segmentation of that microstructure For example, in Figure 2, the yellow and blue points represent the pixels selected to segment the pearlite and graphite microstructure of a gray cast iron Whenever this selection process is finish, the user starts the training of the network used, using for this the data selected After the network training, the system is ready to automatic segments and quantifies the input image The output results

of this procedure are a segmented image, which can be saved for posterior consideration, and the numerical values of the quantification obtained, that can be also saved and exported

to others applications, like the Microsoft Excel

It is important to note that the training of the neural network used is only necessary

to be done once for each type of sample to be analyzed In addition, the SVRNA system lets the pseudocolors adjust of the segmentation obtained, that can be used to improve the

For Peer Review Only

next automatic segmentation results as the network can learn the enhanced done by the user

4 Results and discussion

In this work, experimental results were obtained applying our SVRNA system and the Image Pro-Plus system as well on samples of nodular, malleable and gray cast irons

This turns possible a comparative analysis between these two systems in the automatic quantification of microstructures in images Additionally, it was also used the manual method of quantification on nodular cast iron [23] Among the considered material samples, the nodular cast iron is the only one that has the graphite with a well-defined morphology and so permits the consideration of the manual method of quantification In all cases, the graphite is associated with the black color in the original image, while the pearlite or ferrite to the gray color

In Table 1, are shown the results obtained using the SVRNA and Image Pro-Plus systems in the case of the nodular cast iron It should be noticed that the results obtained using these two systems are very similar, presenting a minimum difference of 2.15% and a maximum of 7.1%, for samples 14 and 9, respectively Moreover, it can be noticed that the SVRNA system presents an average of graphite equal to 12.09% and pearlite or ferrite to 87.91% and that Image Pro-Plus presents 15.93% and 84.07% of graphite and pearlite or ferrite, respectively, being the difference between the two systems equal to 3.84%

For Peer Review Only

Figures 3a), b) and c), present an original image of nodular cast iron, and the images resulting of the segmentation process accomplished using the Image Pro-Plus and SVRNA systems, respectively These images are the ones that present the major difference in the results obtained by the two systems By other hand, images of Figures 3d), e) and f), are the ones that present the minor difference obtained by the same systems

It should be noticed that the segmentation done using our SVRNA system is more accurate than the one obtained using the Image Pro-Plus system, since part of the pearlite or ferrite was erroneously segmented and classified as graphite by Image Pro-Plus, which not occurred with our SVRNA system This fact is the main justification for the existing difference in the results obtained by those two systems

In Table 2, are presented the average results obtained using the manual quantification, SVRNA and Image Pro-Plus system, being the results obtained by the manual method easily distinguished from the ones obtained by the two computational systems This occurs because the manual quantification tends to be a procedure extremely tedious and fastidious, being so propitious for generating errors Moreover, to do this kind

of manual quantification on nodular cast iron, was necessary to analyze 206 points in the same sample, using a reticulated mesh with 25 intersections, in order to obtain an error of 5% [24] The average results obtained using the manual procedure were 24.13% for graphite and 75.87% for ferrite or pearlite

For the gray cast iron it was not possible to do the manual quantification, as the graphite presents forms like ribbings (lodes) what turns this quantification procedure inadequate However, the systems SVRNA and Image Pro-Plus can adequately segment and quantify the constituent of this type of iron as well

For Peer Review Only

In Table 3, are presented the results of the quantification done using the SVRNA and Image Pro-Plus systems These are the most similar results obtained by these systems

on all analyzed samples, presenting a minimum difference of 0.1% and maximum of 6.44%, for samples 1 and 11, respectively Moreover, it can be noticed that our SVRNA system presents an average of graphite equal to 10.28% and of pearlite or ferrite to 89.62%, and that Image Pro-Plus system presents 12.09% and 87.91% of graphite and of pearlite or ferrite, respectively, being the difference between the two systems equal to 1.71%

Figures 4 a), b) and c), present the original image of the grey cast iron, and the resulting images of the segmentation done using the Image Pro-Plus and the SVRNA systems, respectively These images are the ones that present the minor difference between the results obtained by these systems On other hand, images of Figures 4 d), e) and f), are the ones that present the major difference between the same results

Analyzing the results obtained, it can be concluded that the segmentations carried through using the SVRNA and Image Pro-Plus systems are sufficiently similar for sample

1, being that difference insignificant Relatively to sample 11, the segmentation carried through by visual inspection is very similar to the ones obtained using the SVRNA and Image Pro-Plus systems; however, the quantification obtained by these systems presents a difference of 6.44%

The experimental results obtained using the SVRNA and Image Pro-Plus systems

on samples of nodular and gray cast iron are very similar, as the associated images present good quality and so its constituents present forms well defined and with well distinct gray tonalities

For Peer Review Only

If SVRNA and Image Pro-Plus systems are applied on images of malleable cast irons with good quality, the obtained results will be also very similar, as the graphite presents the same tonality that has in the other irons, being just different in its morphology, that, in this case, has a form similar to flakes Thus, it was decided to apply the SVRNA and Image Pro-Plus systems in images of malleable cast iron of low quality in which the ash levels of graphite are similar to the one of pearlite or ferrite This permits to evaluate the efficiency and efficacy of the SVRNA and Image Pro-Plus systems when used in more adverse conditions

In Table 4, are presented the results obtained by the quantification done using the SVRNA and Image Pro-Plus systems on malleable cast iron These results are the most dissimilar for the studied samples, presenting a minimal difference of 3.55% and maximum

of 15.16% for samples 18 and 4, respectively Moreover, it can be noticed that our SVRNA system presents a media of graphite equal to 14.98% and of pearlite or ferrite to 85.02% and the Image Pro-Plus system presents 22.89% and 77.11% of graphite and pearlite or ferrite, respectively, being the involved difference between the systems equal to 7.91%

Figures 5 a), b) and c), present an original image of malleable cast iron and the images resulting from the segmentation done using the Image Pro-Plus and SVRNA systems, respectively These images present the minor difference verified between the results obtained by the used computational systems In other hand, the images of Figures 5 d), e) and f), present the major difference

Verifying the results obtained, it can be noticed that the segmentations carried through using the SVRNA and Image Pro-Plus systems are considerable distinct on sample

4 For sample 18, the segmentations obtained are also distinct, as the Image Pro-Plus

For Peer Review Only

system segments great part of the pearlite or of ferrite, which are due because of the bad quality of the original input image However, that error is not verify when using our SVRNA system that segments correctly the graphite from the other constituents

In the results obtained by the Image Pro-Plus system on samples of nodular, gray and, in particular, malleable casting iron, we can verify the great difficulty of this system to segment the graphite adequately when the background of the input image is not uniform

However, our SVRNA system does not present that deficiency and so it is able to get results effectively

Our SVRNA system shows to be a versatile tool and easy to use in automatic segmentation of material microstructures presented in images, even when the analyzed images are of low quality Moreover, when compared with the Image Pro-Plus system, our SVRNA system needs less time for the quantification of the structures presented, once in the Image Pro-Plus system the adjustment of the pseudocolors to the microstructures presented needs to be made manually and carefully The referred manual procedure can be verified in Figure 6, in which the circle drawn in yellow detaches the manual adjustment of the threshold value that should be used

4.1 SVRNA on other microstructures

Our SVRNA system can also be applied in other microstructures of metallic materials, such as the Image Pro-Plus system

In Figure 7a), a microphotography of AISI 1020 steel is showed in its original form, representing the black color the pearlite grain and the white color the ferrite The

For Peer Review Only

segmentation obtained using our SVRNA system is present in Figure 7b), corresponding the green color to pearlite grain and yellow color to ferrite As we can see from these images, our SVRNA system segmented adequately the microstructures of pearlite and ferrite in this AISI 1020 steel sample

In Figure 8a), a microphotography of AISI 1045 steel is showed in its original form, representing the black color the pearlite grain and the white color the ferrite After the segmentation done using the SVRNA system, the red color corresponds to pearlite grain and the black color to ferrite, Figure 8b) As we can see in these images, our SVRNA system segmented adequately the microstructures of pearlite and ferrite presented in this AISI 1045 steel sample

Another possible application of our SVRNA system is in the segmentation of inclusions and its quantifications in metallic materials Albuquerque et al [25] consider the effect of non-metallic inclusions in AISI 4140 steels on the HAZ toughness in the welding technique of two-layer The results obtained showed that the amount and the length of the non-metallic inclusions could influence negatively the toughness of the HAZ To measure the amount of inclusions presents in the materials from images, was used the SVRNA system as well, that show to be an adequate tool for this kind application also

Two of the results obtained using our SVRNA system in the segmentation and quantification of inclusions are showed in Figure 9

For Peer Review Only

5 Conclusions

This work describes a new computational system, denominated by SVRNA, developed for the segmentation and quantification of the constituents of metallic materials from images, which is based on artificial neural networks

After the presentation of our SVRNA system, was done a comparative analysis on the experimental results obtained using the SVRNA and Image Pro-Plus systems in the segmentation and quantification of microstructures in images of nodular, malleable and gray cast iron, being our SVRNA the system that had the best results

Using several metallographics experiments, in different metal samples, we can conclude that our SVRNA system can be successfully used in applications of the Materials Science field; in particular, for the segmentation and quantification of material microstructures in images Relatively to the Image Pro-Plus system, the SVRNA system presents as main advantages the considerable reduction in the necessary quantification time

as well as segmentations of higher quality

Additional advantages can be observed in the utilization of our SVRNA system, as the easiness of use, the robustness of the neural network employed to noise presented in the analyzed images, mainly, because of the optic distortions or illumination irregularities due

to the image acquisition process Thus, we can confirm that, when compared with the Image Pro-Plus system, our SVRNA system guarantees results that are more satisfactory and in reduced time From the analysis of the experimental results obtained, we can also

For Peer Review Only

conclude that our SVRNA system is sufficient efficient for the degree of significance usually adopted in this domain

In resume, we can conclude that our SVRNA system is able to be used by researchers, engineers, specialists and others of Materials Science field, being so an adequate option to optimize the process of segmentation the microstructures of materials in images and obtain sufficient accurate quantification results

6 Acknowledgments

To Federal Center of Technological Education of Ceará - CEFET CE, for the support given during the accomplishment of this work, in particular to Mechanical Testing Laboratory for the carrying out of the metallographic tests, and to Teleinformation Laboratory, where occurred the optimization of the images, considered in the analyses reported in this work, using for that techniques of Image Processing and Computational Vision systems The authors would like to thank also to CAPES for their financial support

References

[1] F P C Souza, A Susin, SIAV – An automatic vehicle identification system, XIII

Brazilian Conference on Automática (2000), pp 1377-1380

For Peer Review Only

[2] B Acha, C Serrano, Image classification based on color and texture analysis, IWISPA

2000 PROGRAM (2000)

[3] F.Van Der Heidjen, Image based measurement systems object recognition and

parameter estimation, England - John Wiley & Sons Inc, 1994

[4] D Plaut, S Nowlan, G E Hinton, Experiments on learning by backpropagation,

Computer Science Department Carnegie – Mellon University, Technical Report CMU-CS (1986), pp 86-126

[5] H K D H Bhadeshia, Neural networks in materials science, ISIJ International 39

(1999), pp 966-979

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

and engineering, Tata McGraw-Hill Publishing Company Ltd., India, 2006

[7] L Miaoquan, X Liu, A Xiong, X Li, Microstructural evolution and modelling of the

hot compression of a TC6 titanium alloy, Materials Characterization 49 (2003), pp

203–209

[8] L Miaoquan, X Liu, A Xiong, X Li, An adaptive prediction model of grain size for

the forging of Ti-6Al-4V alloy based on the fuzzy neural networks, Journal of

Materials Processing Technology 123 (2002), pp 377-381

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

welding using an intelligent system, The International Journal of Advanced

Manufacturing Technology 22 (2003)

[10] J Kusiak, R Kuziak, Modelling of microstructure and mechanical properties of steel

using the artificial neural network, Journal of Materials Processing Technology 127

For Peer Review Only

[11] X Li, L Miaoquan, Microstructure evolution model based on deformation

mechanism of titanium alloy in hot forming, Transactions of non ferrous metals

society of China 15 (2005), pp 749-753

[12] R Biernacki, J KozUowski, D Myszka, M Perzyk, Prediction of properties of

austempered ductile iron assisted by artificial neural network, Materials Science

(Medžiagotyra) 12 (2006), pp 11-15

[13] A Abdelhay, Application of artificial neural networks to predict the carbon content

and the grain size for carbon steels, Egyptian Journal of Solids 25 (2002), pp 229

– 243

[14] O Wang, J Lai, D Sun, Artificial neural network models for predicting flow stress

and microstructure evolution of a hydrogenized titanium alloy, Key Engineering

Materials (2007), pp 541 – 544

[15] Media Cybernetics, Image-Pro Plus - application notes, Silver Spring: Media

Cybernetics, www.mediacy.com/action.htm (2007)

[16] T Chow, Neural networks and computing, World Scientific Pub, USA, 2007

[17] S Haykin, Neural networks: a comprehensive foundation, Macmillian College

Publishing Company Inc, USA, 1994

[18] W Callister, Materials science and engineering: an introduction, John Wiley & Sons

Inc, USA, 2006

[19] V Chiaverini, Tratamentos térmicos das ligas ferrosas, Associação Brasileira de

Metais, 2ª edição, São Paulo, Brazil, 1987

[20] D Raabe, F Roters, F Frédéric Barlat, L Chen, Continuum scale simulation of

engineering materials, Wiley InterScience Newsletter, 2005

For Peer Review Only

[21] W Wang, Engineering alloys: properties and applications Marcel Dekker Inc, USA,

2007

[22] D Zhang, T C Leia,, Z Zhang,, J Ouyang, The effects of heat treatment on

microstructure and erosion properties of laser surface-clad Ni-base alloy, Surface

and Coatings Technology 115 (1999), pp 176 – 183

[23] V H C Albuquerque, P C Cortez, A R Alexandria, Image segmentation system for

quantification of microstructures in metals using artificial neural networks, Revista

Matéria 12 (2007), pp 394-407

[24] A Seabra, Correlação das propriedades mecânicas dos aços com a microestrutura.

Lisboa - Memória LNEC nº 522, 1979

[25] V H C Albuquerque, C C Silva, C.R.O Moura, W M Aguiar, J.P Farias,

Effect of base metal characteristics on the success of welding of the AISI 4140 steel without post welding heat treatment, XXXIV National Welding Congress (2007)

For Peer Review Only

Figure Captions

Figure 1: Phase diagram for iron-carbon iron (from [18])

Figure 2: Interface of our SVRNA system: on the left, an original image; on the right, the correspondent segmented image for two microstructures

Figure 3: Two original images of nodular cast iron, a) and d); Resulting segmentations using the Image Pro-plus, b) and e), and SVRNA systems, c) and f)

Figure 4: Original images of gray cast iron, a) and d); resulting images of the segmentations done using the Image Pro-Plus, b) and e), and SVRNA systems, c) and f)

Figure 5: Original image of malleable cast iron, a) and d); resultant images of the segmentations done using the Image Pro-Plus, b) and e), and SVRNA systems, c) and f)

Figure 6: Manual adjustment process of the pseudocolors that should be done in the Image Pro-Plus system

Figure 7: Original image of the AISI 1020 steel, a), and after the segmentation done using the SVRNA system, b)

For Peer Review Only

Figure 8: Original image of AISI 1045 steel, a), and after the segmentation done using the SVRNA system, b)

Figure 9: Original images of inclusions, a) and c), segmented images resulting using the SVRNA system, b) and d)

For Peer Review Only

Figure 1: Phase diagram for iron-carbon iron (from [18])

For Peer Review Only

Figure 2: Interface of our SVRNA system: on the left, an original image; on the right, the

correspondent segmented image for two microstructures

For Peer Review Only

Figure 3: Two original images of nodular cast iron, a) and d); Resulting segmentations

using the Image Pro-plus, b) and e), and SVRNA systems, c) and f)

For Peer Review Only

Figure 3: Two original images of nodular cast iron, a) and d); Resulting segmentations

using the Image Pro-plus, b) and e), and SVRNA systems, c) and f)

For Peer Review Only

Figure 3: Two original images of nodular cast iron, a) and d); Resulting segmentations

using the Image Pro-plus, b) and e), and SVRNA systems, c) and f)

For Peer Review Only

Figure 3: Two original images of nodular cast iron, a) and d); Resulting segmentations

using the Image Pro-plus, b) and e), and SVRNA systems, c) and f)