Analysis of local appearance-based face recognition: Effects of feature selection and feature normalization

In this paper, the effects of feature selection and feature normalization to the performance of a local appearance based face recognition scheme are presented. From the local features that are extracted using block-based discrete cosine transform, three feature sets are derived. These local feature vectors are normalized in two different ways; by making them unit norm and by dividing each coefficient to its standard deviation that is learned from the training set. The input test face images are then classified using four different distance measures: L1 norm, L2 norm, cosine angle and covariance between feature vectors. Extensive experiments have been conducted on the AR and CMU PIE face databases. The experimental results show the importance of using appropriate feature sets and doing normalization on the feature vector.