Online Training for Face Recognition System Using Improved PCA

The variation in illumination is one of the main challenging problem for face recognition. It has been proven that in face recognition, differences caused by illumination variations are more significant than differences between individuals. Recognizing face reliably across changes in pose and illumination using PCA has proved to be a much harder problem because eigenfaces method comparing the intensity of the pixel. To solve this problem, this research proposes an online face recognition system using improved PCA for a service robot in indoor environment based on stereo vision. Tested images are improved by generating random values for varying the intensity of face images. A program for online training is also developed where the tested images are captured real-time from camera. Varying illumination in tested images will increase the accuracy using ITS face database which its accuracy is 95.5 %, higher than ATT face database's as 95.4% and Indian face database's as 72%. The results from this experiment are still evaluated to be improved in the future.


INTRODUCTION
Ability for face recognition and real interaction with a user is one important issue for developing vision-based service robots. Since face tracking and face recognition are essential functions for a service robot, many researchers developes face-tracking mechanism for the robot (Yang M., 2002) and face recognition system for service robot ( Budiharto, W., 2010). The objective of this chapter is to propose an online training for face recognition system using improved principal component analysis (PCA) which is implemented to a service robot in a dynamic environment using stereo vision.
Variation in illumination is a challenging problem for face recognition. It has been proved that differences caused by illumination variations are more significant than ones between individuals (Adini et al., 1997). Recognizing face reliably across changes in pose and illumination using PCA is proved to be a much harder problem because of eigenfaces method compared to the intensity of the pixel. To solve this problem, we have improved the training images by generating random values for varying the intensity of face images.
We have proposed an online face recognition system using PCA. This model is very important because it can be implemented to service robots so that they are able to automatically learn and recognize the customers. Several experiments using three poses images (front, left and right) of each person and given training images with varying illumination improves the success rate for recognition. Our proposed method is successfully implemented to a service robot called Srikandi III in our laboratory.

Improved Face Recognition Using PCA
Face is our primary focus of attention in developing a vision-based service robot. Unfortunatelly, developing a computational model of face recognition is quite difficult, because faces are complex and multidimensional. Modelling of face images can be based on a statistical model like principal component analysis (PCA) (Turk and Pentland, 1991 ) and linear discriminant analysis (LDA) (Etemad & Chellappa, 1997;Belhumeur et.al, 1997), as well as on a physical model on the assumption of certain surface reflectance properties, such as Lambertian surface (Zoue et al., 2007). Linear discriminant analysis (LDA) is a method for finding such a linear combination of variables which best separates two or more classes. Constrasting their PCA which encoded information in an orthogonal linear space, LDA which is also known as fischerfaces method encodes discriminatory information in a linear separable space of which bases are not necessary orthogonal. However, the LDA result is mostly used as a part of a linear classifier (Zhao et al., 1998).
PCA is a standard statistical method for feature extraction by reducing the dimension of input data by a linear projection that maximizes the scatter of all project samples. The scheme is based on an information theory approach that decomposes face images into a small set of characteristic feature images called "eigenfaces", a principal component of the initial training set of face images. Recognition is performed by projecting a new image into the subspace spanned by the eigenfaces called "face space", and then classifying the face by comparing its position in face space with the positions of known individuals. PCA based approaches typically include two phases: training and classification. In the training phase, an eignespace is established from the training samples using PCA and the training face images are mapped to the eigenspace for classification. In the classification phase, an input face is projected to the same eignespace and classified by an appropriate classifier (Turk & Pentland, 1991 ). Let a face image I (x,y) be a two-dimensional N by N array of (8-bit)  For storing user's training faces, we propose a simple face table database (face_id, name, date_registered, image_file1, image_file2, image_file3) as shown in Figure 3:

METHOD
We have developed a method for online training for face recognition system. The training images with extention .pgm is stored in a directory. In online mode, the program will store the face images to files. In training mode, the training images will be executed for training. In testing mode, the input image from camera will be compared with the training images. The algorithm is shown below: Algorithm 1. Online training for Face recognition system using imporved PCA.

Begin
Call We have developed a vision-based service robot called Srikandi III with the ability to perform face recognition and avoid people as moving obstacles. This wheeled robot is the next generation of Srikandi II (Budiharto, 2010). The prototype of service Robot Srikandi III utilizing a low cost stereo Minoru 3D camera is shown in Figure 4:

RESULTS AND DISCUSSION
We have identified the effect of varying illumination to the recognition accuracy for our database called ITS face database as shown in Table 1. Results shows that by giving enough training images with variation of illumination generated randomly, the success rate of face recognition will be improved. If the recoqnition test does not apply illumination varying, and enough training images are available, PCA can recognize people's faces with success rate up to 100%. But if varying illumination is applied and the training images are not sufficient (for example six tested images with only six training images) the success rate only 50%. However, if the number of training images are increased, the success rate can reach 100%. Based on Table 1, too many tested images with varied illumination will decrease the success rate from 100% to 91.60% because the training images are not adequate to recognize the variation of illumination from tested images. Therefore, the best way to increase the succes rate with many tested images with illumination variation is have adequate number of training images. We also evaluate the result of our proposed face recognition system and compared with ATT and Indian face database using Face Recognition Evaluator developed by Matlab. Each face database consists of ten sets of people's face. Each set of ITS face database consists of 3 poses (front, left, right side) and varied with illumination. ATT face database consists of nine differential facial expression and small occlusion (by glass) without variation of illumination . The Indian face database consists of eleven pose orientations without variation of illumination and the size of each image is too small than one of ITS and ATT face database. The success rate comparison among the three face databases is shown in Figure 5. From the figure it is clearly noticed that ITS database have highest accuracy than ATT and Indian face database when the illumination of the tested images is varied. The accuracy using PCA in ITS face database as much as 95.5 %, higher than ATT face database as 95.4% and IFD face database as 72%.