Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face...

4
The International Arab Conference on Information Technology (ACIT’2013) Biometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi Faculty of Engineering and Information Technologies, International Burch University,Bosnia and Herzegovina [email protected] , [email protected] Abstract:Artificial methods as pattern recognition, machine learning, and artificial neural network are used for facial recognition. In this paper we present a LMT where we have used fallowing methods: extract skin color, then converted into histogram values for performing LMT - logistic model tree. LMT is standard decision tree. The nodes of terminal are replaced with logic regression function. The accuracy of face recognition is 91%. We used histogram numerical values for performing supervised leering tasks for the prediction of nominal classes and numerical values. Here is presented LMT which adopts the idea of classification problems using logistic regression. We have run algorithm from our dataset. Keywords:Face recognition, LMT, biometric security. 1. Introduction This study has aim to show the results of logistic model tree and its behavior in the human facial recognition. The prior work has been done related to biometric recognition [1]. Facial recognition using tree method has shown significant accuracy of 91%. Due the rapid technology improvements, much more research is obtained for improving facial recognition. Since, facial recognition is very complex issues due its high complex features as: facial shapes, skin color, facial behavior, etc., this brings us to very interesting study area. In this paper we will present the results of facial recognition using LMT. Comparing with previous research in [2, 3], substantial improvements have been seen by applying LMT methods regarding PCA in [2, 3]. Many of prior work on biometry related to facial recognition have been examined. There are more approaches and ways to determine research in this field in order to retrieve the best results, concerned under various conditions. Such conditions are: facial expression, skin detection, shadow, light feature on camera etc. Therefore, PCA (Principle Component Analysis), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis) are the most common used in facial recognition, in here we present the LMT which is also prominent technique for supervised learning tasks and adopts the idea of classification using logistic regression, instead of linear regression The results were retrieved from the IBU (International Burch University) facial database. The performance was measure among 40 persons. This paper consists of: 1. Introductions 2. Background and LMT 3. Empirical Results and Discussion 4. Conclusion 2. Background and LMT For facial recognition we have used logistic model tree – logistic regression classifier to evaluate the recognition results and accuracy. Logistic regression approach can be described as function of : →, or(|), where Y is the classifier which describes the discrete value. The vector which has any discrete or continues value is described as: =< ... > In general Logistic Model Tree consists of standard decision tree combined with logistic regression functions. The result provided in this paper are the outcome of methodology and experiment performing LMT, shows that LMT is better compared to PCA method [2, 3].The algorithm of LMT is described in Algorithm 1, and is described as [4]: - Growing tree: LMT grows at the root using LogitBoost algorithm. This is used for fitting iterative simple linear regression functions. It uses cross validation. Moreover it includes the criteria when to stop building the tree as well what to do with a different types of attributes [4]. - Building model: Then it splits the data at the root of the tree, where tree is constructed. Splitting can be for numerical or nominal. The tree grows as it sorts the appropriate subsets of the data to the child node and continues to build logistic model tree. - Splitting continues while any of criteria is met [4]. - Pruning: When tree is building us uses CART based pruning [4].

Transcript of Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face...

Page 1: Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi Faculty of Engineering

The International Arab Conference on Information Technology (ACIT’2013)

Biometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi

Faculty of Engineering and Information Technologies, International Burch University,Bosnia and Herzegovina

[email protected] , [email protected]

Abstract:Artificial methods as pattern recognition, machine learning, and artificial neural network are used for facial recognition. In this paper we present a LMT where we have used fallowing methods: extract skin color, then converted into histogram values for performing LMT - logistic model tree. LMT is standard decision tree. The nodes of terminal are replaced with logic regression function. The accuracy of face recognition is 91%. We used histogram numerical values for performing supervised leering tasks for the prediction of nominal classes and numerical values. Here is presented LMT which adopts the idea of classification problems using logistic regression. We have run algorithm from our dataset. Keywords:Face recognition, LMT, biometric security.

1. Introduction This study has aim to show the results of logistic model tree and its behavior in the human facial recognition. The prior work has been done related to biometric recognition [1]. Facial recognition using tree method has shown significant accuracy of 91%. Due the rapid technology improvements, much more research is obtained for improving facial recognition. Since, facial recognition is very complex issues due its high complex features as: facial shapes, skin color, facial behavior, etc., this brings us to very interesting study area. In this paper we will present the results of facial recognition using LMT. Comparing with previous research in [2, 3], substantial improvements have been seen by applying LMT methods regarding PCA in [2, 3]. Many of prior work on biometry related to facial recognition have been examined. There are more approaches and ways to determine research in this field in order to retrieve the best results, concerned under various conditions. Such conditions are: facial expression, skin detection, shadow, light feature on camera etc. Therefore, PCA (Principle Component Analysis), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis) are the most common used in facial recognition, in here we present the LMT which is also prominent technique for supervised learning tasks and adopts the idea of classification using logistic regression, instead of linear regression The results were retrieved from the IBU (International Burch University) facial database. The performance was measure among 40 persons. This paper consists of:

1. Introductions 2. Background and LMT 3. Empirical Results and Discussion 4. Conclusion

2. Background and LMT

For facial recognition we have used logistic model tree – logistic regression classifier to evaluate the recognition results and accuracy. Logistic regression approach can be described as function of푓: 푋 → 푌, or푃(푌|푋), where Y is the classifier which describes the discrete value. The vector which has any discrete or continues value is described as: 푋 =< 푋 . . . 푋 > In general Logistic Model Tree consists of standard decision tree combined with logistic regression functions. The result provided in this paper are the outcome of methodology and experiment performing LMT, shows that LMT is better compared to PCA method [2, 3].The algorithm of LMT is described in Algorithm 1, and is described as [4]:

- Growing tree: LMT grows at the root using LogitBoost algorithm. This is used for fitting iterative simple linear regression functions. It uses cross validation. Moreover it includes the criteria when to stop building the tree as well what to do with a different types of attributes [4].

- Building model: Then it splits the data at the root of the tree, where tree is constructed. Splitting can be for numerical or nominal. The tree grows as it sorts the appropriate subsets of the data to the child node and continues to build logistic model tree.

- Splitting continues while any of criteria is met [4].

- Pruning: When tree is building us uses CART based pruning [4].

Page 2: Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi Faculty of Engineering

Figure 1. Model of facial recognition LMT(example){ root = new Node() alpha = getCARTAlpha(example) root.buildTree(examples, null) root.CARTprune(alpha) } buildTree(examples, initialLinearModels){ numIterations = CV_Iterations(examples, initalLinearModels) initLogistBoost(initialLinearModels) linearModels = copyOf(initialLinearModeles,examples) for i = 1...numIterations logitBoostIterations(linearModels,examples) split = findSplit(examples) localExamples = split.splitExamples(examples) sons = new Nodes[split.numSubsets()] for s = 1...sons.length soons.buildTree(localExamples[s],nodeModels) } CV_Iterations(examples, initalLinearModesl{ for fold = 1...5 initLogitBoost(initialLinearModels) //split into training / test train = trainCV(fold) linearModels = copyOf(initialLinearModels) for i = 1...200 logitBoostIterations(linearModels, train) logErrors[i] += error(test) numIterations = findBestIterations (logErrors) return numIterations }

Algorithm 1. Pseudo code of LMT [4]

In Algorithm 1, has shown the building of pseudo code of LMT algorithm. Logistic Model Tree combines logistic regression models with tree induction. It is applicable model for tree classification problem. Actually, we can see that pseudo code and LMT method constructs the tree from the training data examples. The primary it calls getCARTAlpha to cross validates parameters. CARTPrunde implements pruning scheme. The method buildTree grows by logistic model tree and recursively splits the instance space. Argument initialLinearModel contains simple linear regression function which fits LogisticBoot at higher level. The method initLogitBoost initializes the probabilities of weights for the LogitBoost algorithm like that already have had fitted with regression function initialLinearModel.

3. Experimental Results and Discussion The evaluation of LMT was conducted among 40 persons in the IBU database. The experiment consists of following steps as shown in Figure 1:

1. Capture Images 2. Skin color detection 3. RGB to Gray 4. Histogram 5. LMT 6. Classification

Figure 2. Facial database sample The presented results of LMT, shows better performance than C4.5 which did not shown enough significant results that would be relevant for considering it to present it in this paper. Cross – validation accuracy (CVA) Eq. (1) is the average of k individual accuracy measures

CVA = ∑ 퐴 (1)

Accuracy we measure by calculating true positive (TP) Eq. (2) and false positive (FP) Eq. (3) rate that are used for evaluating the performance. TruePositiveRate(TPR) = × 100% (2) FalsePositiveRate(FPR) = × 100% (3) Accuracy Eq. (4) is used to represent the overall measure and is represented as:

Page 3: Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi Faculty of Engineering

88%90%92%94%96%98%

Logistic Model Tree

Logistic Model Tree

Accuracy = × 100% (4) One more measure which we have performed in here is F-measure Eq.(5). F-measure is described as: F −measure = × 100% (5) The ROC is used to show the evaluation of discrimination of the classifier’s ability. The curve of ROC denotes the performance of classifier. In [5] is presented the calculation of ROC. The ROC is described on plot with representation of sensitivity vs. 1 – specificity. On Y – axis are plotted TPR – true positive rate and on X – axis are plotted FPR – false negative rate.

Table 1: Testing results

Table 2: ROC, F-measure and TP results

The image parameters were applied with LMT. The parameter consist from histogram values which were retrieved from facial images, skin color detection. During the testing phase, LMT were selected to according k-fold procedure. It consisted of 10 k-fold cross validations. The accuracy of using LMT is 91%. Comparing to study for testing facial recognition for mobile phone detection in [2, 3] was 88%. The training phase consisted of 20 poses for 40 persons. It consisted of males and females. The experiment included 800 images. In Table 1 - 2 are described results of LMT. Figure 3 shows graphical representation of results presented in Table 2.

Figure 3. Graphical Representation

4. Conclusion In here we have shown LMT behavior in artificial model of facial recognition. This is the continuation of [2, 3], where for the future we would like to come up with solution which model would fit best in software solution for facial recognition application. Moreover, we would look up for the model, that would be have the best solution in meeting various conditions. The evolution of LMT has shown 91% accuracy. In this paper we have introduces the possibility of applying LMT in facial recognition. Such method includes the LogitBoost algorithm. Our research in this paper shows that LMT produces model which is more accurate than perhaps C4.5. A further research will be to try to achieve a better accuracy by researching in trying to handle missing values. Even, database did not consist of even more images, it show significant result and LMT behavior for encouraging in continuation of further research procedure in this scope. Such research in future should provide better approaches for computer biometric security in the applications. References: [1] A. Subasi, "Classification of EMG signals using

PSO optimized SVM for diagnosis of neuromuscular disorder," Computers in Biology and Medicine, no. 43, pp. 576-586, 2013.

[2] T. Hastie, R. Tibshirani and J. Friedman, The Elements of Statistical Learning Data Mining, Inferecne and Prediction, Stanford, California: Springer, 2008.

[3] E. Kremic, A. Subasi and K. Hajdarevic, "Face Recognition Implementation for Client Server Mobile Arhitecture," in 34th International Conference on Information Technology Interfaces, Dubrovnik, Croatia, 2012.

[4] E. Kremic and A. Subasi, "The Implementation of Face Security for Authentication Implemented on Mobile Phone," in The 12th International Arab Conference on Information Technology, Riyadh, Saudi Arabia, 2011.

[5] H. Khalajzadeh, M. Mansouri and M. Teshnehlab, "Face Recognition using Convolution Neural Network and Simple Logistic Classifier," in Online Conference on Soft Computing Industry, 2012.

[6] A. Kumar and D. Zhang, "Biometric Recognition using Feature Selection and Combination".

Correctly Classified Instance

Incorrectly Classified Instance

Time Taken to Build Model

LMT 91.00 8.00 3

ROC F-measure TP

LMT 97% 91% 91%

Page 4: Biometry: Face Recognition Applying Logisticacit2k.org/ACIT/2013Proceedings/192.pdfBiometry: Face Recognition Applying Logistic Emir Kremic, Abdulhamit Subasi Faculty of Engineering

[7] N. Landwehr, M. Hall and E. Frank, "Logistic Model Tree," in 14th European Conference on Machine Learning, 2003.

Emir Kremic graduated from University of Buckingham, UK in 2008. He took his M.Sc. degree from International Burch University, Sarajevo, Bosnia and Herzegovina in 2011. He is pursuing his PhD from the International Burch University. He is reviewer in International Journal of Computer and Electrical Engineering. His areas of interests are: mobile security, biometric security using facial and video recognition, pattern recognition, machine learning. Emir Kremic has served as Judge at International Program Olympiad for IT at International Burch University. As well he was serving as volunteer as reviwer on IEEE conferences. Abdulhamit Subasi graduated from Hacettepe University in 1990. He took his M.Sc. degree from Middle East Technical University in 1993, and his Ph.D. degree from Sakarya University in 2001, all in Electrical and Electronics Engineering. In 2006 he was senior researcher at Georgia Institute of Technology, School of Electrical and Computer Engineering, USA. From 2001 until 2008 he has been an Assistant Professor in the Department of Electrical and Electronics engineering at Kahramanmaras Sutcu Imam University. Since 2008 he is appointed as Associate Professor of Information Technology and Dean of Engineering Faculty at International Burch University. His areas of interest are data mining, machine learning, pattern recognition, biomedical signal processing, computer networks and security. He has worked on several projects related with biomedical signal processing and pattern classification. Dr. Subasi has served (or is currently serving) as a program organizing committee member of the national and international conferences. He is editorial board member of several scientific journals. Moreover, he is voluntarily serving as a technical publication reviewer for many respected scientific journals and conferences. He has lots of published journal and conference papers on his research areas.