| dc.description.abstract | The human face is always used in the application of access control systems, identity authentication systems, surveillance systems and security systems. Some researh on face recognition has been proposed with a variety of methods, such as Eigen Faces, Fisher’s Linear Discriminant, Bayesian Networks, Neural Networks,Hidden Markov Model, and so on. But all the above mentioned methods have limitations, such as dealing with the size of the image and lighting and learning time are too long. In a study to obtain an alternative method for face recognition, Le TH and Bui L (2011) conducted research model of face recognition with Two- DimensionalPrincipal Component Analysis (2DPCA) and Support Vector Machine (SVM) using AT&T face image data. The results of these studies that the accuracy of face recognition use 2DPCA and SVM was 97.3% and PCA-SVM was 95.2%. Then in 2010, Bi-2DPCA uses as alternative method of face recognition performed by Yang J et al. (2010). In a study comparing methods of Bi-2DPCA, 2DPCA and PCA. The conclusion that the use of Bi-2DPCA generating and computational accuracy was better than 2DPCA and PCA Based on research conducted by Le TH and Bui L (2011) and Yang J et al. (2010), conducted research for the development of models of face recognition using dimension reduction technique Bi-2DPCA and Support Vector Machine (SVM) as pattern recognition. 1DPCA and 2DPCA dimension reduction technique are used as comparison measure dimensions and computing time than using Bi-2DPCA technique. To help SVM in recognizing human faces, SVM use linear, polynomial and radial basis function (RBF) kernel function. When using RBF kernel, used for the RBF kernel, the parameter for sigma is 2-15 up to 23 and the value of C starting from 2-5 up to 214 which have been used in a study conducted by Le TH and Bui L (2011). Facial image data used are taken from the ORL Face Database AT&T Laboratory, which contain 40 faces, each face with 10 images, and total of image 400 pieces. Validation is done by Leave-One-Out (LOO) cross validation, so that for each pair of extraction and classifier method is 10 trials. One against all method used to help SVM in classifying the ORL face data consists of 40 classes, to reduce the computational time. From the results of tests on 36 models of face recognition , Bi-2DPCA85% with the help of a linear kernel generates the highest level of face recognition accuracy of 94.25 %, and the lowest accuracy rate of 24.50 % was obtained with 1DPCA-kernel polynomial ordo 3 . While the lowest computational time of 15.34 seconds when component is used by 85 % computation time and a maximum of 252.68 seconds when component is 95 %. There are several things that could be concluded that: pattern accuracy on each kernel feature extraction are the same. This means that there is no relationship between the feature extraction used by the kernel to the SVM. The second is, in this study SVM cannot classify large dimension data. From the results of tests on 36 models of face recognition, Bi2DPCA85% with the help of a linear kernel generates the highest level of recognition accuracy of 94.25% face, whereas the lowest accuracy rate of 24.50% was obtained with the kernel polynomial method 1DPCA order 3. While the lowest computational time of 15,34 seconds when the component is used by 85% computation time and a maximum of 252.68 seconds when the component is 95%. There are several things that could be concluded that: pattern accuracy on each kernel feature extraction are the same. This means that there is no relationship between the feature extraction used by the kernel to the SVM. The second is, in this study SVM cannot classify large data dimension . The larger the dimension reduction is performed, the higher the accuracy value obtained. Third is, the higher the component does not automatically increase accuracy. The conclusion of the study is Bi-2DPCA method with SVM is a good model of face recognition where the highest accuracy with the lowest time than 2DPCA-SVM and SVM- 1DPCA. | en |
