Biometric recognition using example data to predict the identity of people. This is the niche of machine learning. Machine learning is computer programming that detect certain pattern from example data to predict output of future unknown input. In face recognition, the input is face image while the output is identity of input images. Compared to other biometric information, face image is easier and more flexible to acquire even when the person is unaware of being scanned.

 

There are two key steps in face recognition: feature extraction and classification. In this paper, Principal Component Analysis plays the key role as feature extractor. Main idea of Principal Component Analysis is to find eigen vectors that best represent the distribution of all face images in dataset. This technique also reduces a large dataset to lower dimension. Principal Component Analysis calculate the covariance matrix of the face vectors and find its eigen vectors. Eigen vectors corresponding to m largest eigen values are chosen to be the projection matrix. This projection matrix is used to compute the m-dimensional final matrix by calculate the product of the transpose of the projection matrix and matrix data.

 

Learning Vector Quantization method is used to classify Labeled Faces in the Wild-a (LFW-a) dataset. Learning Vector Quantization network has an input layer, a competitive layer and a linear layer. The competitive layer aims to classify input vectors to subclasses, meanwhile the second linear layer aims to combine subclasses into the target classes. Those layers have one neuron for each class, meaning that if competitive layer has p neurons, it can learn up to p classes. Then these are combined by the linear layer to produce q target classes where p is always larger than q. In 2017, Li and Lin compared performance of Support Vector Machine, Principal Component Analysis + Support Vector Machine, and Histogram of Oriented Gradient + Support Vector Machine on LFW dataset. The highest recognition accuracy is under 70% for those three algorithms. The authors of this research are still developing the algorithm. Using Principal Component Analysis + Learning Vector Quantization classificator and adding fuzzy kernel to the network, the recognition accuracy is expected to be above 80%.