Proceedings of the 2 nd International Conference on Computer Engineering & Mathematical Sciences (ICCEMS 2013)

Transcription

1 The nd International Conference on Computer Engineering & Mathematical Sciences (ICCEMS 0) Proceedings of the nd International Conference on Computer Engineering & Mathematical Sciences (ICCEMS 0) 5-6 December 0, Kuala Lumpur, Malaysia

2 The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) Recognizing Occluded and Illuminated Faces using ICA Algorithm Ebtesam N. Abdullah AlShemmary, Zaid Abdi Alkareem Alyasseri,,Ibrahim Venkat, Zoha Pourebtehaj and MarwahHaithamKhattab, Information Technology Researchand Development Center,University of Kufa, An Najaf, Iraq School of Computer Sciences,Universiti Sains Malaysia, Penang, Malaysia University of Mosul, Mosul, Iraq {dr.alshemmary, Abstract: Face recognition system compares the tested face with the various training faces reserved in the database with efficient success rate. The best matching of the tested face with the training faces is an important task. In this paper, how to recognize a face is presented, two different analysis algorithms are included in the evaluation system, Eigenface and ICA. The local dataset used in this paper are pre-processed using statistical standard methods.preprocessing software which is provided by the Colorado State University (CSU) Face Identification Evaluation System Version 5.0 under Unix Shell scripts was written using ANSII C code. Independent Component Analysis algorithm ICA is written using Matlab for face recognition implementation.this paper explains how the faces, having some variations like facial expressions and viewing conditions w.r.t the original faces reserved in the database, are detected with improved accuracy and success rate. Finally the result shows several graphs by Matlab and Excel. Keywords: Image Processing,Face Recognition System, ICA algorithm, Eigenface.. Introduction During the past three decades, extensive research has been conducted on automatically recognizing the identity of individuals from their facial images. In spite of the existence of alternative technologies such as fingerprint and iris recognition, human face remains one of the most popular cues for identity recognition in biometrics. Face recognition possesses the non-intrusive nature and are often effective without the participant's cooperation or knowledge. It makes a good compromise between performance reliability and social acceptance and well balances security and privacy. Other biometric methods do not possess these advantages. For instance, fingerprint recognition methods require the subjects to cooperate in making explicit physical contact with the sensor surface []. Similarly, iris recognition methods require the subjects to cooperate in placing their eyescarefully relative to the camera. Nowadays, automatic face recognition has become one of the most popular area of research in computer vision and one of the most successful applications of image analysis and understanding []. Face recognition system is a biometric technique which is used in law enforcement agencies, personal identification systems, and security purposes. This computer application automatically identifies a person from a digital image [].Face recognition can be done by several algorithms such as PCA, ICA, LDA, EP, EBGM, AMM, SVM, HMM and etc. []. Principal component analysis (PCA) is a popular unsupervised statistical method to find useful image 57 representations.some of the most successful representations for face recognition,such as eigenfaces [4], holons [5], and local featureanalysis [6] are based on PCA. In a task such as face recognition,much of the important information may be containedin the high-order relationships among the image pixels, andthus, it is important to investigate whether generalizations ofthe PCA which are sensitive to high-order relationships, not justsecond-order relationships, are advantageous. Independentcomponent analysis (ICA) [7] is one such generalization. An algorithm for performing ICA has been proposed. In this paper we dothe task using ICA algorithm and eigenvectors.the approach transform face images into a small set of characteristic feature images, called "eigenface", which are the principal component of the initial training set of face images [8], then we apply the ICA (Independent component analysis) which is a statistical and computational technique for revealing hidden factors that underlie sets of random variables, measurements, or signals [9], the first step for a face recognition system is to recognize a human face and extract it from the rest of the scene. Next, the system measures nodal points on the face, such as the eyes and mouth; these nodal points are used to remove the background of the image and keep the statistical feature (distinct face). The rest of this paper is organized as follows. Section provides the necessary related work to face recognition. Section describes enrollment images for the system (.), and preprocessing the local data set images (.), and image recognition system (.). Section 4 shows experimental results and discussion. Section 5 concludes with practical recommendations.. Related Work The development of face recognition over the past years allows an organization into three types of recognition algorithms, namely frontal, profile, and view tolerant recognition, depending on the kind of images and the recognition algorithms. This section gives an overview on the major human facerecognition techniques that apply mostly to frontal faces,advantages and disadvantages of each method are also given.the methods considered are eigenfaces (eigenfeatures), neuralnetworks, dynamic link architecture, hidden Markov model,geometrical feature matching, and template matching. Theapproaches are analyzed in terms of the facial representationsthey used [0].

3 Recognition Part Preprocessing Part The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) Bell and Sejnowski [, ] developed an algorithm from the point of view of optimal information transfer in neural networks with sigmoidal transfer functions. This algorithm has proven successful for separating randomly mixed auditory signals (the cocktail party problem), and for separating electroencephalogram (EEG) signals and functional magnetic resonance imaging (fmri) signals. The performance of ICA depends on the task, the algorithm used to approximate ICA, and the number of subspace dimensions retained. Even more confusing, there are two very different applications of ICA to face recognition. ICA can be applied so as to treat images as random variables and pixels as observations, or to treat pixels as random variables and images as observations. In keeping with [, 4], we refer to these two alternatives as ICA architecture I and architecture II, respectively. Alternatively, ICAarchitecture I produce spatially localized features that are only influenced by small parts of the image. It has beenargued that this will produce better object recognition, since it implements recognition by parts. If localizedfeatures are indeed superior, ICA architecture I should outperform PCA and ICA architecture II.This paper will show empirically that the choice of subspace projection algorithm depends first and foremost onthe nature of the task. Some tasks, such as facial identity recognition, are holistic and do best with global featurevectors. Other tasks, such as facial action recognition, are local and do better with localized feature vectors. For bothtypes of tasks, ICA can outperform PCA, but only if the ICAarchitecture is selected with respect to the task type(ica architecture I for localized tasks, ICA architecture II for holistic tasks). Furthermore, performance is optimizedif the ICA algorithm is selected based on the architecture (InfoMax for architecture I, FastICA for architecture II).When PCA is used, the choice of subspace distance measure again depends on the task.. Preprocessing The purpose of preprocessing is to remove artifacts from the local dataset images. Preprocessing in this paper has been done by implementingthe following: - Feature Cartesian coordinates: marked using Face Recognition by Independent Component Analysis and implement the program for "markfeatures.m" using MATLAB. For each image mark the eyes and mouth coordinate ( person at the time, to avoid confusing name) (we need to change the directory of "imgdir" and "outputdir" in the source file), the result from this program will be saved on a text file.table show example of the output data. Dataset Images mark the eyes and mouth coordinate for all the images Implement csupreprocessnormalize.c on C Preprocessing Images Implement the ICAScript.m on Matlab using training images folder Implement the final.m on MatLab using testing images folder Result Rank form to : Image has been recognized. Proposed Method The proposed system consists mainly of two parts, which is preprocessing and image recognition. Figure shows the overall system that used in this paper and refers to the system performance too. Figure. Overall system.. Enrollment By using our local dataset images with extension.jpg for face recognition system, we have 4 students' images each student has 8 images (Neutral, Smiling, open mouth, dim, bright, scarf, sunglass, scarf + sunglass ). The data sets have been dividedinto two groups: training folder and testing folder, (as shown in Figure ). The training folder has images (Neutral, Smiling and Open mouth), so its contain 7images, and the testing folder contains five subfolders (folder for Dim light images,folder for Bright light images, folder for Sunglass images,folder 4 for Scarf images and folder 5 for Sunglass + Scarf images), So each subfolder contains 4 images. Figure. The Dataset folder division 58

4 Vector N x The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) Image name Table. Part of the output data from step Right eye Left eye X,Y position X, y position Mouth X, y position Dim Normal Open mouth Light Scarf Smile Figure. Preprocessing Implementation s Sunglas Sunglas s_ scarf Image Convertion and saving:then by using the csufaceideval_5. [5] which usingansii C code under Linux and do as flowing: a- Extention conversion: the extension of images must be converted from.jpg into.pgm by using the command: ls *.jpg awk -F. '{system("convert " $0 " " $ ".pgm")}' We must make sure that the current directory is in the picture file that needs to be converting into.pgm. b- Makefile compilation:compile the makefile in the file"\\user\csufaceideval_5." by type make (e.g.: $ make), we will see an execution file appear in \\user\csufaceideval_5.\bin\x86_64. c- Running: run the file name (csupreprocessnormalize) at user/csufaceideval_5./bin/x86_64 by type./csupreprocessnormalize. d- Execution file option: An option of the execution file will appear, choose the option./csupreprocessnormalize - pgm <output directory><coordinate.txt directory><picture directory>, Replace the <> with your own directory and run the command. e- Testing: check the output directory to see the normalize.pgm file. f- File opening: finally open the.pgm file in the MatLab. Figure shows the implementation after preprocessing phase on our dataset on ANSCII C.Then the aligned faces savedas a new dataset. Original image Normal image Smile image Open mouth image. Image Recognition A template matching problem considered the simplest approach of image recognition.problems arise when performing recognition in a high- dimensional space. Significant improvements can be achieved by first mapping the data into a lower dimensional space.figure 4explains Arc II method. The lower-dimensional space found using idea behind eigenface[6]: - Suppose Γ is an N x vector, corresponding to an NxN face image I. - The idea is to represent Γ (Φ=Γ - mean face) into a lowdimensional space: Φ ˆ mean = w u + w u +... w K u K (K<<N ) N N Image Figure 4. Arch II method The computation of eigenfaces is as follows [4]: Step : obtain face images I, I,..., IM (training faces) (Very important: the face images must be centered and of the same size), as in Figure5. Figure 5. The training faces. Step : represent every image Ii as a vector Γi Step : compute the average face vector Ψ: Step 4: subtract the mean face: 59

5 The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) Step 5: compute the covariance matrix C: Pseudo code PROCEDUR for the proposed method refers how our algorithmswork, see (Appendix ).Figure 6shows how the pcabigfn.m file works.the results for implementing Eigenface are shown in Figure 7. Figure 9.Bright image result using ICA Figure6.The pcabigfn.m file flow work. Figure 0.Scarf image result using ICA Figure 7. The results for Eigenface. 4. Experimental Results and Discussion We have implemented the ICA algorithm using MATLAB on Dell INSPIRON Core Duo.0 Ghz CPU with 4 GB of RAM. We used dataset which has 4 persons, each person has 8 images (natural, bright, dim-light, smile, open mouth,sunglass, scarf and scarf&sunglass). The total time requirements to implement the program about minutes. The ICA results are represented based on the Cumulative Match Characteristic (CMC) where the CMC shows the curve of plots of the probability between :N where N is the number of the persons in the dataset.figures(from 8to4) showsthe implemention of Arch for face recognition using the ICA algorithm on different images such as (Normal, bright, dim-light, smile, open mouth, sunglass, scarf and scarf+sunglass). We find that the resultsof illumination imagesare better than the occlusion images, (see Figures and 4). Figure.Sunglass image result using ICA Figure.Sunglass and scarf image result using ICA Figure.Illumination images result using ICA Figure 8.Dim-light image result using ICA 60

6 The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) ure 4.Occlusion images result using ICA 5. Conclusion and future work Fig In this paper we implemented face recognition system based on the ICA algorithm by using Matlab R00b. We used a local dataset which has 4 students' images each student has 8 images (Neutral, Smiling, open mouth, dim, bright, scarf, sunglass, scarf + sunglass ). The data sets have been dividedinto two groups: training folder and testing folder. The training folder has images (Neutral, Smiling and Open mouth), so its contain 7images, and the testing folder contains five subfolders for (Dim light images, Bright light images, Sunglass images,scarf images and Sunglass + Scarf images), So each subfolder contains 4 images. We find that the srbe erbtsebare in illumination images (Dim light images, Bright light images). There is some limitation in the current system, this limitation has occurred as a result of the presence of some low-lying-resolution images or the images contain noise so we need to do some enhancement as a postpreprocessing. Figure 5 shows some of these images. Figure 5: Show some low resolution images For future work weare looking to reduce the current run time also we suggest to do some image enhancement methods to remove blurring and noising from the original images such as using the harmony search algorithm for image enhancement [7, 8]. 6. Acknowledgement The authors would like to thank Dr. Ibrahim Venkat and all the postgraduate students 0/0 from School of computer science/university Science Malaysia for supporting our work. References [] Ebtesam N. AlShemmary, Fingerprint Image Enhancement and Recognition Algorithms, PhD. Thesis, University of Technology, Baghdad, 4 June, 007. [] [] [4] M. Turk and A. Pentland, Eigenfaces for Recognition, J. CognitiveNeurosci., vol., no., pp. 7 86, 99. [5] G. Cottrell and J. Metcalfe, Face, Gender and Emotion Recognition Using Holons, in Advances in Neural Information Processing Systems, D. Touretzky, Ed. San Mateo, CA: Morgan Kaufmann, 99, vol., pp [6] P. S. Penev and J. J. Atick, Local Feature Analysis: A General Statistical Theory for Object Representation, Network: Comput. Neural Syst., vol. 7, no., pp , 996. [7] P. Comon, Independent Component Analysis A New Concept, Signal Processing, vol. 6, pp. 87 4, 994. [8] Matthew A. Turk and Alex P. Pentland, "Face Recognition Using Eigenfaces", CH98-5/9/0000/0586 IEEE, 99. [9] [0] Rui Huang, Vladimir Pavlovic, and Dimitris N. Metaxas, "A Hybrid Face Recognition Method Using Markov Random Fields", ICPR (), pp , 004. [] A. J. Bell and T. J. Sejnowski, An Information- Maximization Approach to Blind Separation and Blind Deconvolution, Neural Comput., vol. 7, no. 6, pp. 9 59, 995. [] Bell, Anthony J., and Terrence J. Sejnowski, The Independent Components of Natural Scenes are Edge Filters, Vision Res., vol. 7, no., pp. 7 8, 997. [] M. S. Bartlett, "Face Image Analysis by Unsupervised Learning: Kluwer Academic", 00. [4] M. S. Bartlett and T. J. Sejnowski, "Viewpoint Invariant Face Recognition Using Independent Component Analysis and Attractor Networks", in Neural Information Processing Systems - Natural and Synthetic, M. Mozer, M. Jordan, and T. Petsche, Eds. Cambridge, MA: MIT Press, vol. 9, pp. 87-8,997. [5] Beveridge, J.R., et al., "The CSU Face Identification Evaluation System", Machine Vision and Applications, 6(), p.p 8-8, 005. [6] M. Turk and A. Pentland, "Eigenfaces for Recognition", Journal of Cognitive Neuroscience, vol., no., pp. 7-86, 99. [7] Peter T. Higgins, "Introduction to Biometrics", Sep., 005. [8] Zaid Abdi Alkareem, Y. A., et al. "Edge preserving image enhancement via harmony search algorithm." Data Mining and Optimization (DMO), 0 4th Conference on. IEEE, 0. 6

7 The nd International Conference on Computer Engineering and Mathematical Sciences (ICCEMS 0) Appendix Procedure for preprocess. ProcedureconvertImages : this function used to preprocess the images by using the [ProceduregenerateMask which is used to (create mask with ship ellipse so that the matrix contain either or 0 according to the Procedure ellipsefunc which is used to (find the distance between specify pixels in the image and the ellipse ) if the distance larger than 0 put in the matrix else put 0) ], read from file all the information like (filename, eye Lift position x, eye Lift position y, eye Right position x, eye Right position y) and apply this make to that image the result will be image without any background just show the statistical feature.. Masks are no longer read from a file. It is created by scan converting an ellipse.. GenerateMask function. 4. ConvertImages function: this function used to remove the background for the image and keep the statical feature for the image. 5. Prepare file suffix encoding preprocessing settings, blank if not requested. 6. Shift the eye coordinates if necessary. 7. Smooth image borders Eigenface Algorithm. Align training images x, x,, xn. Compute average face u = /N Σ xi. Compute the difference image φi = xi u 4. Compute the covariance matrix (total scatter matrix) ST = (/N) Σ φi φi T = BBT, B=[φ, φ φn] 5. Compute the eigenvectors of the covariance matrix ST 6. Compute training projections a, a... an Testing. Take query image X. Project X into Eigenface space (W = {eigenfaces}) and compute projection ωi = W(X u),. Compare projection ωi with all training N projections ai Procedure for image recognition Step : In loadfacemat.m file we get the aligned and preprocessed training images (we have people images each of them have images for training: Neutral, smiling, open mouth) so we put all these images in a matrix called C. Step : The main program is in ICAscript.m file. So we pass C to this file. Here the function called pcabigfn.m in which we calculate the eigenvectors of each image. The outputs of this function would be V that the eigenvectors are in columns of it, R whose rows are the pca representation of each observation and E containing the eigenvalues. Step : In this phase, the eigenfaces generated from the last step on a figure by using show_ Einginface_ Function.s Step 4: ICA algorithm will be done by applying Arch to the data. As the eigenvectors are in columns of V we put 7 eigenvectors of training images to variable x. Then we run the ICA algorithm on x. For running architecture we should run the function runarch.m. in that function we use the 7 images in variable R put it in x and then again run ICA algorithm on it by calling function runica.m on it. Step 5: To run the ICA algorithm on our matrix of images we have a file names runica.m. In this file assume image gravalues are in rows of x. it will find N independent components, where N is the number of images. Note x gets overwritten. Step 6: We have a function called spherex.m. it spheres the training matrix x to remove first and second order stats from it.we calculate the whitening matrix in this file.and then we start to train the algorithm with these training sets. Step 7: During the training phase we have defined the learning rule in a function called sep96.m. Step 8: In runthisfirst.m we have the codes concerning the ranking of each image.for comparing the training set and testing set we compute the faces by projecting the images into the subspace and measuring the cosine distance between them. We used the nearest neighbor algorithm using cosines as the similarity measure in the file nnclassfn.m We calculate the distance between the testing image and each training image and save the result in a variable called rank. Step 9: Now we run Get_id.m function.we have 5 folders as the testing set. Each of these folders contains 4 images we gave each person an ID starting from to 4. Our training set contains 7 images of 4 persons it means that each person has Pseudo code PROCEDUR 6