Two-step Feature Extraction in a Transform domain for face recognition

Abstract

A face recognition system using an integration of Discrete Cosine Transform (DCT) and Support Vector Machine (SVM) is proposed in this paper. Feature Extraction and Identification are the two main phases of the system. The first phase consists of a preprocessing step, which includes cropping and resizing techniques, followed by DCT coefficient selection and SVM classifier creation. The final outputs contain the DCT coefficients beside several two-input SVM classifiers. A DCT selection algorithm is employed to retain the coefficients which have the maximum variability across each training pose. The data from the nearest, as measured by Euclidean distance, two subjects is used as an input to the SVM classifier. The second phase aims to find the recognition rates based on the Euclidean distance criterion and the output(s) of SVM classifier(s). Four different image databases, namely, ORL, YALE, FERET, and Cropped AR are used to evaluate the system. The proposed system is shown to outperform some of the state of the art systems in terms of the recognition rates.