International Journal of Computing

Research Institute of Intelligent Computer Systems

Ternopil National Economic University

2009, Vol. 8, Issue 1


Contents and abstracts

  1. K. Madani and R. Kh. Sadykhov. Editorial. - p. 5-7.
  2. M. Kamrul Islam. An Improved Architecture for Competitive and Cooperative Neurons (CCNS) in Neural Networks. - p. 8-15.
  3. E. Volna. Forming Evolutionary Design of Neural Networks with Different Nodes. - p. 16-23.
  4. P. C. Hung, S. F. McLoone, R. Farrell. Direct and Indirect Classification of High Frequency LNA Gain Performance – A Comparison between SVMs and MLPs. - p. 24-31.
  5. M. Voiry, K. Madani, V. Amarger, J. Bernier. Data Dimensionality Reduction for Neural Based Classification of Optical Surfaces Defects. - p. 32-42.
  6. I. Budnyk, A. Chebira, K. Madani. Estimating Complexity of Classification Tasks Using Neurocomputers Technology. - p. 43-52.
  7. R. A. Vazquez, H. Sossa. Associative Memories Network for Face Recognition and Object Recognition. - p. 53-60.
  8. Y. Kurylyak, I. Paliy, A. Sachenko, A. Chohra, K. Madani. Face Detection on Grayscale and Color Images Using Combined Cascade of Classifiers. - p. 61-71.
  9. H. Kim, J. K. Tan, S. Ishikawa, T. Shinomiya. Automatic Detection of Spinal Deformity Based on Statistical Features from the Moire Topographic Images. - p. 72-78.
  10. W. Bellil, C. Ben Amar, A. M. Alimi. A Survey on Wavelet Network, Multi Library Wavelet Network Training, 1D-2D Function Approximation and a New Image Compression Method. - p. 79-86.
  11. R. Kh. Sadykhov, V. V. Ganchenko, L. P. Podenok. Fuzzy Clustering Methods in Multispectral Satellite Image Segmentation. - p. 87-94.

EDITORIAL “Artificial Neural Networks and Intelligent Information Processing”

Guest Editors: Kurosh Madani 1) and Rauf Kh. Sadykhov 2)

1) Prof. Dr. Kurosh Madani
PARIS-EST / PARIS 12 – Val de Marne University
Senart-FB Institute of Technology / LISSI Lab.
IUT Senart – Bat. A, Av. Pierre Point
F-77127 Lieusaint, France
Email: madani@univ-paris12.fr
2) Prof. Dr. Rauf Kh. Sadykhov
Computer Systems Department
University of Informatics and Radioelecrtronics
6 P. Brovka st, Minsk, Belarus.
Email: rsadykhov@bsuir.by

New applicative and technological challenges emanating from industrial, socioeconomic or environmental needs, appeared in recent decades, have opened new dilemmas which have decisively highlighted limitations of conventional computational science and issued techniques. Recent borders’ contraction between biological and computational sciences, especially the latest developments in bio-inspired artificial systems over the last decade, may play a central role in designing adequate solutions to these new challenging dilemmas. The fantastic ever-increasing intellectual dynamics created around bio-inspired Artificial Intelligence and related topics (as Artificial Neural Networks, Humanoid Robotics, Ambient Intelligence, etc…), uphold by escalating interest of both confirmed and young researchers on this relatively juvenile science, generates a reach multidisciplinary synergy between a large number of scientific communities making conceivable a forthcoming emergence of viable solutions to aforementioned real-world applicative and technological challenges.

This special issue of the International Scientific Journal of Computing includes a selection of papers presented at the Third International Workshop on Artificial Neural Networks and Intelligent Information Processing (ANNIIP), which was held in Angers, France, May 9-12th, 2007. Since 2005, the ANNIIP international workshop takes part in the aforementioned appealing dynamics within the frame of the prestigious IFAC/IEEE ICINCO International Conference. The objective of this workshop is to convene a set of relevant recent works focusing Bio-inspired Artificial Intelligence related fields and applications by offering a privileged space to overhaul and exchange the knowledge about further theoretical advances, new experimental discoveries and novel technological improvements in this promising area.

ANNIIP points toward the choice of a relatively restricted number of papers. Such philosophy is motivated on the one hand by the premeditated desire to give a large space to exchanges and discussions during the workshop, and on the other hand by the strong principle of the presentation of each accepted article by its authors. Conformably to our philosophy and reaching those objectives, the idea of the present Special Issue has been motivated by our desire to devote an additional “scientific space” to a number of selected papers of ANNIIP 2007 in order to complete and to extend the presented works.

Articles composing this Special Issue are extended versions of already accepted papers and presented during ANNIIP 2007. That’s why we have attached a special attention on reviewing process of this Special Issue. In fact, our first attention has been our premeditated desire to involve totally new experts not belonging to ANNIIP 2007 Program Committee. We would like to reedit our gratitude to all those experts for the valuable work that they have accomplished in total freedom and independency.

The papers selected for this special issue reflect the above-proclaimed deliberate philosophy and the variety of research presented during the workshop. They epitomize miscellany of involved topics and the diversity of techniques connecting Artificial Neural Networks, genetic algorithms, evolutionary computation, machine learning and experts’ hybridization.

The following two papers are devoted to Machine learning mechanisms, design and architecture of Artificial Neural Networks.

The paper “An Improved Architecture for Competitive and Cooperative Neurons (CCNS) in Neural Networks” by K. Islam describes competition and cooperation in neural processing dealing especially with the storing ability of the memory model. The paper proposes a new architecture for competitive and cooperative neurons improving and increasing the storing ability.

The paper “Forming Evolutionary Design of Neural Networks With Different Nodes” by E. Volna presents the design of a neuron-evolution based system. The presented system combines populations of neurons in order to form the fully connected multilayer feed-forward neural networks with fixed architecture. The article shows the transfer function’s important impact on obtained artificial neural network’s architecture and its significant impact on artificial neural network’s performance. The efficiency of described method has been evaluated on pattern recognition as well as on alphabet coding problem.

Complex data and information classifications as well as models’ mixing strategies (e.g. intelligent systems) have traditionally been one of the main pillars of ANNIIP. The three following papers reflect three significant aspects of the aforementioned areas. The first one is the paper “Direct and Indirect Classification of High Frequency LNA Gain Performance – A Comparison between SVMs and MLPs” by P. C. Hung, S. F. McLoone and R. Farrell, which deals with direct and indirect ANN based classification. An indirect Multilayer Perceptron (MLP) and direct support vector machine (SVM) classification strategy are considered, evaluated and compared. The evaluation is performed considering the challenging problem of low noise amplifiers (LNA) design and their high-frequency performances in functional testing. A novel testing strategy using machine learning classifiers to predict highfrequency LNA gain performance by combining information from several lower frequency measurements has been elaborated. The reported case study shows the proposed technique’s potential in extending the operating frequency range by 20% at 2 GHz and 42% at 1.4 GHz.

The second key aspect relating classification skills is data dimensionality reduction. The paper “Data Dimensionality Reduction for Neural Based Classification of Optical Surfaces Defects” by M. Voiry, K. Madani, V. Amarger and J. Bernier is concerned with this purpose. Within the frame of a real industrial application, dealing with high-tech optical devices production, the paper compares different techniques which permit dimensionality reduction and evaluate their impact on ANN based classification tasks performances. Principal Component Analysis (PCA), Self Organizing Maps (SOM), Curvilinear Component Analysis (CCA) and Curvilinear Distance Analysis (CDA) issued techniques have been used in order to reduce the dimensionality of the initial feature space. The paper shows that using CDA the global defect detection and correct classification performances reach 95% defecting performances obtained when using raw data (not reduced feature space).

Finally, the third key aspect coping with complex data classification is “classification task’s complexity estimation which is the purpose of the paper “Estimating Complexity of Classification Tasks Using Neuro-Computers Technology” by I. Budnyk, A. Chebira and K. Madani. This paper presents an alternative approach for estimating classification tasks’ complexity. The frame of the presented work deals with the construction of a selforganizing neural tree-like structure, following the “divide and rule” paradigm. A new approach using IBM Zero Instruction Set Computer (ZISC-036) neuro-processor is described, evaluated and applied to a range of the different classification tasks dealing with real-world classifications dilemmas.

If applications of Artificial Neural Networks have been, over the last decade, an ever increasing subject in numerous conferences relating these bioinspired models (including ANNIIP), their innovative applications continue to be a noteworthy part of the workshop. During ANNIIP 2007, this component has taken a thrilled space. This can be appreciated in the five selected papers of this Special Issue. Two among them focus face recognition, a rising concern in nowadays public security domain. The three others deal with further skills of pattern and image recognition.

The paper “Associative Memories Network for Face Recognition and Object Recognition” by R. A. Vazquez and H. Sossa is devoted to a network of associative memories (AMs) to recall a collection of patterns. The accuracy of the proposed AMs is evaluated on the one hand on an objects’ recognition application and the other hand on a faces’ recognition one. First the all, the benchmarks are split into several collections and then this collections are used to train the network of AMs. During training an image of a collection is associated with the rest of the images belonging to the same collection. Once trained the network we expected to recover a collection of images by using as an input pattern any image belonging to the collection.

The paper “Face Detection on Grayscale and Color Images Using Combined Cascade of Classifiers” by Y. Kurylyak, I. Paliy, A. Sachenko, A. Chohra and K. Madani is another emissary representative of how to deal with the face recognition applications combining several classifiers. The paper describes improved face detection methods for grayscale and color images using a cascade of classifiers and skin color segmentation. The proposed multi-classifiers face authentication method allows achieving one of the best detection rates on available benchmarks with a high processing speed making the proposed approach suitable for a video flow processing. The paper also shows that the use of a mixture of color spaces is more efficient for the skin color segmentation than the consideration of only one color space.

In the paper “Automatic Detection of Spinal Deformity Based on Statistical Features from the Moire Topographic Images” by H. Kim, J. K. Tan, S. Ishikawa and T. Shinomiya, a new automatic classification method for the spinal deformity detection. The focused application concerns the biomedical area relating a disease mainly suffered by teenagers during their growth stage particularly from elementary school to middle school. The proposed technique combines an automatic moire image based Regions of interest’s (ROI) extractor and Artificial Neural Network based classifier. Several Artificial Neural Network based classifiers have been evaluated and compared. The obtained results reach the classification rate of 85% making the proposed concept a suitable candidate for designing computer aided spinal deformity detection tool.

The paper “A Survey on Wavelet Network, Multi- Library Wavelet Network Training, 1D-2D Function Approximation and a New Image Compression Method” by W. Bellil, C. Ben Amar and A. M. Alimi focuses Wavelet Network and its Multi library training within the frame of image compression application. The proposed approach takes advantage from an optimization procedure using a wavelet functions’ library. The efficiency of color images compression is increased comparing to a number of conventional techniques. The loss in processing speed can be largely corrected by decreasing the number of wavelets in hidden layer without decreasing considerably the compression ratio.

Finally, the paper “Fuzzy Clustering Methods in Multi-spectral Satellite Images Segmentation” by R. Kh. Sadykhov, V. V. Ganchenko and L. P. Podenok, presents the core concept Fuzzy Clustering (FC) based satellite images segmentation. The three FC methods (C-means, Gustafson-Kessel, and Gath-Geva algorithms), with and without preliminary nonlinear filtering, testify that segmentation using fuzzy clustering methods provides good-looking discrimination of land cover types that occurs in the complex cases. Applied on Landsat multi-spectral images, the obtain results show the effectiveness of proposed concept in segmentation of wetland, water-meadow, and bush areas which remain awkward points in satellite images’ analysis. The quality of segmented images was approved by experts on the basis of land-based expedition data.

Before ending the editorial, it is important to remind that scientific relevance and technical quality of a collective issue emerge from quality of its contributors: those who contribute by the high quality of their manuscripts. We would like to express again our acknowledgements to contributors of all selected papers: You are the central reason of the nobles of this Special Issue.

It is also essential to be reminiscent that frequently, creative dynamics is the result of fruitful human contacts within a same scientific field or the consequence of humans’ interactions from different scientific communities and since 2004, the date of the its first edition, the ICINCO multi-conference has been an outstanding bench of such creative synergies. For that, we would like to express our particular gratitude to Prof. Joaquim Filipe, ICINCO 2007 Conference’s Chair, for his faith in young science of “Bio-inspired Artificial Intelligence” and for his reliance on including once more the ANNIIP workshop within his valuable conference. We would like also be thankful to Prof. Anatoly Sachenko for devoting this privileged space of his journal to ANNIIP 2007.

Finally, we would like to apology for the somehow tardy reviewing processing, which have been, nevertheless, the condition to guarantee the freedom and total independency of Reviewing Board in accomplishment of their valuable expertise.

Top


AN IMPROVED ARCHITECTURE FOR COMPETITIVE AND
COOPERATIVE NEURONS (CCNS) IN NEURAL NETWORKS

M. Kamrul Islam

School of Computing,
Queen’s University,
Kingston, K7L 3N6, ON, Canada
islam@cs.queensu.ca

In neural networks, the associative memory is one in which applying some input pattern leads to the response of a corresponding stored pattern. During the learning phase the memory is fed with a number of input vectors and in the recall phase when some known input is presented to it, the network recalls and reproduces the output vector. Here, we improve and increase the storing ability of the memory model proposed in [1]. We show that there are certain instances where their algorithm can not produce the desired performance by retrieving exactly the correct vector. That is, in their algorithm, a number of output vectors can become activated from the stimulus of an input vector while the desired output is just a single vector. Our proposed solution overcomes this and uniquely determines the output vector as some input vector is applied. Thus we provide a more general scenario of this neural network memory model consisting of Competitive Cooperative Neurons (CCNs).

Top


FORMING EVOLUTIONARY DESIGN OF NEURAL NETWORKS
WITH DIFFERENT NODES

Eva Volna

University of Ostrava,
30th Dubna st. 22,
701 03 Ostrava, Czech Republic
e-mail: eva.volna@osu.cz, http://www.osu.cz

Evolution in artificial neural networks (e.g. neuroevolution) searches through the space of behaviours for a network that performs well at a given task. Here is presented a neuroevolution system evolving populations of neurons that are combined to form the fully connected multilayer feedforward neural network with fixed architecture. In this article, the transfer function has been shown to be an important part of architecture of the artificial neural network and have significant impact on an artificial neural network’s performance. In order to test the efficiency of described method, we applied it to the pattern recognition problem and to the alphabet coding problem.

Top


DIRECT AND INDIRECT CLASSIFICATION OF HIGH FREQUENCY LNA
GAIN PERFORMANCE – A COMPARISON BETWEEN SVMS AND MLPS

Peter C. Hung, Seán F. McLoone, Ronan Farrell

Institute of Microelectronics and Wireless Systems,
Department of Electronic Engineering,
National University of Ireland Maynooth,
Maynooth, Co. Kildare, Ireland,
{phung, sean.mcloone, rfarrell}@eeng.nuim.ie,
http://imws.eeng.nuim.ie/

The task of determining low noise amplifier (LNA) high-frequency performance in functional testing is as challenging as designing the circuit itself due to the difficulties associated with bringing high frequency signals offchip. One possible strategy for circumventing these difficulties is to inferentially estimate the high frequency performance measures from measurements taken at lower, more accessible, frequencies. This paper investigates the effectiveness of this strategy for classifying the high frequency gain of the amplifier, a key LNA performance parameter. An indirect Multilayer Perceptron (MLP) and direct support vector machine (SVM) classification strategy are considered. Extensive Monte-Carlo simulations show promising results with both methods, with the indirect MLP classifiers marginally outperforming SVMs.

Top


DATA DIMENSIONALITY REDUCTION FOR NEURAL BASED
CLASSIFICATION OF OPTICAL SURFACES DEFECTS

Matthieu Voiry 1&2) , Kurosh Madani 1), Véronique Amarger 1), Joël Bernier 2)

1) Image, Signal and Intelligent Systems Laboratory (LISSI / EA 3956),
Senart Institute of Technology,
University PARIS XII, Av. Pierre Point,
F-77127 Lieusaint, France,
{voiry ; madani ; amarger}@univ-paris12.fr,
http://www.univ-paris12.fr/
2) SAGEM REOSC
Avenue de la Tour Maury, Saint Pierre du Perray, 91280, France
{mathieu.voiry or joel.bernier}@sagem.com

A major step for high-quality optical surfaces faults diagnosis concerns scratches and digs defects characterization in products. This challenging operation is very important since it is directly linked with the produced optical component’s quality. A classification phase is mandatory to complete optical devices diagnosis since a number of correctable defects are usually present beside the potential “abiding” ones. Unfortunately relevant data extracted from raw image during defects detection phase are high dimensional. This can have harmful effect on the behaviors of artificial neural networks which are suitable to perform such a challenging classification. Reducing data dimension to a smaller value can decrease the problems related to high dimensionality. In this paper we compare different techniques which permit dimensionality reduction and evaluate their impact on classification tasks performances.

Top


ESTIMATING COMPLEXITY OF CLASSIFICATION TASKS USING
NEUROCOMPUTERS TECHNOLOGY

Ivan Budnyk, Abdennasser Chebira, Kurosh Madani

Images, Signals and Intelligent Systems Laboratory (LISSI / EA 3956)
PARIS 12 – Val de Marne University, Senart-Fontainebleau Institute of Technology,
Bat. A, Av. Pierre Point, F-77127 Lieusaint, France,
{ivan.budnyk, chebira, madani}@univ-paris12.fr,
http://lissi.univ-paris12.fr

This paper presents an alternative approach for estimating task complexity. Construction of a self-organizing neural tree structure, following the paradigm “divide and rule”, requires knowledge about task complexity. Our aim is to determine complexity indicator function and to hallmark its’ main properties. A new approach uses IBM © Zero Instruction Set Computer (ZISC-036 ®) and applies for a range of the different classification tasks.

Top


ASSOCIATIVE MEMORIES NETWORK FOR FACE RECOGNITION AND
OBJECT RECOGNITION

Roberto A. Vazquez 1), Humberto Sossa 2)

1) Center for Computing Research (CIC-IPN), Av. Juan de Dios Batiz s/n Col. Nueva Industrial Vallejo, CP. 07738
Mexico City, Mexico, ravem@ipn.mx
2) Center for Computing Research (CIC-IPN), Av. Juan de Dios Batiz s/n Colonia Nueva Industrial Vallejo CP. 07738
Mexico City, Mexico, hsossa@cic.ipn.mx

An associative memory (AM) is a special kind of neural network that allows associating an output pattern with an input pattern. Some problems require associating several output patterns with a unique pattern. Classical associative and neural models cannot solve this simple task and less if these patterns are complex images, for example faces. In this paper a network of AMs to recall a collection of patterns is proposed. The accuracy of the proposal is tested with two benchmarks. One is composed by 20 objects and the other is composed by 20 images of 15 different people faces. First the all, the benchmarks are split into several collections and then this collections are used to train the network of AMs. During training an image of a collection is associated with the rest of the images belonging to the same collection. Once trained the network we expected to recover a collection of images by using as an input pattern any image belonging to the collection.

Top


FACE DETECTION ON GRAYSCALE AND COLOR IMAGES USING
COMBINED CASCADE OF CLASSIFIERS

Yuriy Kurylyak 1), Ihor Paliy 1), Anatoly Sachenko 1), Amine Chohra 2), Kurosh Madani 2)

1) Research Institute of Intelligent Computer Systems,
Ternopil National Economic University,
3 Peremoga Square, 46004, Ternopil, Ukraine
{yuk, ipl, as}@tneu.edu.ua
2) Images, Signals and Intelligent Systems Laboratory (LISSI / EA 3956),
PARIS XII University, Senart-FB Institute of Technology
Av. Pierre Point, Bat. A, F-77127, Lieusaint, France
{chohra, madani}@univ-paris 12.fr

The paper describes improved face detection methods for grayscale and color images using the combined cascade of classifiers and skin color segmentation. The combined cascade with proposed face candidates’ verification method allows achieving one of the best detection rates on CMU test set and a high processing speed suitable for a video flow processing. It’s also shown that the mixture of color spaces is more efficient during the skin color segmentation than the application of one color space. A lot of experiments are made to choose rational parameters for the developed face detection system in order to improve the detection rate, false positives’ number and system’s speed.

Top


AUTOMATIC DETECTION OF SPINAL DEFORMITY BASED ON
STATISTICAL FEATURES FROM THE MOIRE TOPOGRAPHIC IMAGES

Hyoungseop Kim 1), Joo Kooi Tan 1), Seiji Ishikawa 1), Takashi Shinomiya 2)

1) Department of Control Engineering, Kyushu Institute of Technology,
1-1, Sensui, Tobata, Kitakyushu 804-8550,
Japan, kim@cntl.kyutech.ac.jp
2) Nikon Co. LTD., Japan

Spinal deformity is one of a disease mainly suffered by teenagers during their growth stage particularly from element school to middle school. There are many different causes of abnormal spinal curves, but all of them are unknown. To find the spinal deformity in early stage, orthopedists have traditionally performed on children a painless examination called a forward bending test in mass screening of school. But this test is neither objective nor reproductive, and the inspection takes much time when applied to medical examination in schools. To solve this problem, a moire method has been proposed which takes moire topographic images of human backs and checks symmetry/asymmetry of their moire patterns. In this paper, we propose a method for automatic judgment of spinal deformity which is obtained moire topographic images based on statistical features on the moire image. Statistical feature of asymmetry degrees are applied to train employing the classifier such as Artificial Neural Network, Support Vector Machine, Self-Organization Map and AdaBoost.

Top


A SURVEY ON WAVELET NETWORK, MULTI LIBRARY WAVELET
NETWORK TRAINING, 1D-2D FUNCTION APPROXIMATION AND A NEW
IMAGE COMPRESSION METHOD

Wajdi Bellil 1), Chokri Ben Amar 2), Adel M.Alimi 3)

1) Faculty of sciences, University of Gafsa, City Zarroug, Gafsa, Tunisia wajdi.bellil@ieee.org
2) Department of Electrical Engineering, University of Sfax, Tunisia Chokri.benamar@ieee.org
3) Department of Electrical Engineering, University of Sfax, Tunisia Adel.Alimi@ieee.org

This paper presents an original architecture of Wavelet Neural Network (WNN) based on multi Wavelets activation function and uses a selection method to determine a set of best wavelets whose centers and dilation parameters are used as initial values for subsequent training library WNN for color image compression and coding which consists to transform an RGB image into Luminance-Chrominance space and then segment the luminance in a set of m blocks n by n pixels. These blocks should be transferred row by row (1D input vector) to the input of our wavelet network. Every input vector will be considered as unknown functional mapping and then it will be approximated by the network.

Top


FUZZY CLUSTERING METHODS IN MULTISPECTRAL SATELLITE
IMAGE SEGMENTATION

Rauf Kh. Sadykhov 1), Valentin V. Ganchenko 1), Leonid P. Podenok 2)

1) Computer Systems Department, Belarusian State University of Informatics and Radioelecrtronics,
6 P. Brovka st, Minsk, Belarus
rsadykhov@bsuir.by, ganchenko@lsi-bas-net.by, http://www.bsuir.by
2) Laboratory of System Identification,
United Institute of Informatics Problems,
National Academy of Sciences of Belarus,
6 Surganov st, Minsk, Belarus
podenok@lsi-bas-net.by, http://uiip.bas-net.by

Segmentation method for subject processing the multi-spectral satellite images based on fuzzy clustering and preliminary non-linear filtering is represented. Three fuzzy clustering algorithms, namely Fuzzy C-means, Gustafson- Kessel, and Gath-Geva have been utilized. The experimental results obtained using these algorithms with and without preliminary nonlinear filtering to segment multi-spectral Landsat images have approved that segmentation based on fuzzy clustering provides good-looking discrimination of different land cover types. Implementations of Fuzzy Cmeans, Gustafson-Kessel, and Gath-Geva algorithms have got linear computational complexity depending on initial cluster amount and image size for single iteration step. They assume internal parallel implementation. The preliminary processing of source channels with nonlinear filter provides more clear cluster discrimination and has as a consequence more clear segment outlining…

Top