Skip to Main Content
Contourlet transform (CT) is a new efficient image representation which identifies two key features of an image that improves over the separable 2-D wavelet transform, namely directionality and anisotropy. In this paper, a method based on contourlet transform and weighted similarity measure (WSM) for face recognition is proposed. Low frequency and the directional high frequency subbands coefficients can be produced by contourlet transformation on face images. For feature extraction, low-frequency coefficients are divided into a few sub-blocks. All of the means and standard deviations of each sub-block constitute low frequency characteristic vectors. On the other hand, the histogram graphs of directional high frequency subband coefficients can be fitted with generalized Gaussian density (GGD) model. The similarity of low-frequency characteristic vectors is measured by Euclidean distance, and that of the high frequency components is measured by Kullback-Leibler (K-L) distance. The WSM is implemented by computing the weighted average of these two kinds of distances. The experimental results show that weighted similarity measure for contourlet-based face recognition can achieve higher recognition rates.