Spectral domain dominant feature extraction algorithm for face and palm-print recognition

Abstract

Automatic face and palm-print recognition have widespread applications in security, authentication, surveillance, and criminal identification. The task of recognition is based on extracting unique characteristics that exist in face and palm-print images. Different environmental factors, such as nonlinear lighting variations, scaling and rotation, and physiological aspects, such as facial expressions and palm ridges, may significantly affect these characteristics. Moreover, different images of a particular person may vary largely, while images of different persons may not necessarily vary significantly. Hence, obtaining a significant feature space with respect to the spatial variation that exists in face or palm-print images, although crucial, becomes a challenging task. In this thesis, an efficient feature extraction algorithm is proposed for face and palm-print recognition based on exploiting local spatial variations in spectral domain. The feature extraction is performed in transform domain, where there exists a scope of utilizing different popular spectral-domain transforms. Reducing the dimension of the feature space is another important task in order to ensure lower computational complexity. Instead of using the entire face or palm-print image as a whole, for feature extraction, certain local zones with higher information content are selected, which not only provides reduced computational burden but also confines our region of interest leaving out spatial redundancies. In view of selecting such high-informative local zones, an entropy based information measure is proposed. The task of feature extraction is carried out within those local zones using discrete Fourier transform, discrete cosine transform and discrete wavelet transform. Moreover, a dominant feature selection algorithm is developed to obtain a significantly discriminative feature space, which also provides a reduction in feature dimension. The merits of extracted local dominant features are analyzed using confidence measures, such as intra-class compactness and inter-class separation. The recognition task is carried out utilizing the extracted features in some distance based classifiers. The recognition performance is evaluated on several standard face and palm-print databases and it has been found that the proposed method provides high recognition accuracy even for images affected due to shift, scaling and nonlinear lighting variations, compared to that obtained by some recent algorithms.