A Multi-scale Threshold Local Binary Pattern In Quaternion Wavelet Transform Domain For Face Recognition

The valid representation of face image is the basis of face recognition, there arealready a lot of face image representation methods, most of them are based onstatistical learning, which inevitably leads to the algorithm trained by a large numberof samples. In this paper, a novel face recognition method based on a multi-scalethreshold local binary pattern in quaternion wavelet transform domain is proposed.The method is not based on statistical learning, which naturally avoids these problems,and has outstanding performance in the case of a single sample.The Quaternion Wavelet Transform (QWT) is a new multi-scale analysis tool forgeometric image features. It is a near shift-invariant tight frame representation whosecoefficients support a magnitude and three phases and can be computed efficiently,amplitude phase information increases the robustness of the algorithm to expressionsand illumination compared with wavelet coefficients. The LBP operator is one of thebest performing texture descriptors and it has been widely used in various applicationsfor texture extraction. However, LBP is influenced by noise, A/D transformation erroreasily. What’s worse, the principal information, such as the shape and position of eyes,nose, mouth and son on, and non-principal information coming from the skin line,noise, facial expression and A/D transformation error usually, are treated equally. Inthis paper, The Threshold LBP is proposed which get principal information and ignorenon-principal information. The combination of QWT and Threshold Local BinaryPattern, not only has the multi-scale multi-resolution the characteristics of the QWT,but also has the texture extraction ability of the threshold local binary patterns indifferent sensitivity. These quaternion magnitude and phases are combined, and thenmulti-scale Threshold LBP features are extracted. We transform distance of amagnitude and three phases to a order of magnitude and achieve better result andchi-square distance is for classify.The proposed approach is compared with several popular algorithms in standardFERET96database, and these experiments show that the influence of differentparameters. The experiment result on FERET dataset confirms the effectiveness ofthis method.