Curvelet texture based face recognition using Principal Component Analysis

A vital issue for face recognition is to represent a face image by effective and efficient features. To-date a numerous feature extraction techniques have been proposed in the literature. Among them, content based image retrieval (CBIR) using curvelet transform captures accurate texture features to represent the image. In this paper, we propose a novel face recognition method that uses curvelet texture features for face representation. Features are computed by low order statistics like mean and standard deviation of transformed face images. Since the spectral domain of curvelet has no hole or overlap, there is no loss of frequency information in face images. Moveover, such feature representation has considerably low dimension. Thus, computation within the face-space becomes easier. Furthermore, the dimension of features is independent of face image resolution. As a result, it can support face images of different resolution as input. To build the classifier, we apply PCA on the concatenated feature representation of subdivisions. We test our system with 4 and 5 levels of scales of curvelet transform. We also experiment by dividing the face image into different number of sub-divisions on three standard databases. The experimental results confirm that curvelet texture features achieve satisfactory performance for face recognition.