Mixed Nonorthogonal Transforms Representation for Face Recognition

Abstract

An alternative face recognition system that additively combines two-dimensional discrete wavelet transform (2D-DWT) coefficients and two-dimensional discrete cosine transform (2D-DCT) coefficients for image feature extraction is proposed. Each training pose is represented by superimposing the dominant coefficients from the two domains taking into account the nonorthogonality of the coefficients in one domain with respect to the coefficients in the other domain. The recognition system is tested with three publicly available databases, namely ORL, YALE, and FERET. As shown in the sample results, the proposed system significantly reduces the required storage size, a desirable property for big data and when computing resources are limited, while maintaining the accuracy of recognition rates when compared with the 2D-DCT, the 2D-DWT, and the successive 2D-DWT/2D-DCT techniques. In addition, the computational complexity in the testing phase is comparable with that of recently reported techniques.