Illumination Invariant Face Recognition Based On Ordinal Features

Illumination variation is one of the most important problems in face recognition. Intensity distribution of the image is changed because of illumination variations. When intensity information is used to recognize face, it would cause inner-class differences (i.e., images of one person in different illumination conditions) are worse than between-class differences (i.e., images of different people in the same illumination condition). This restricts the applications of face recognition technology. State-of-the-art methods can be roughly divided into three groups including illumination subspace based methods, illumination normalization and illumination invariant extraction.This thesis proposes an improved face recognition algorithm based on ordinal features, which aim at acting as illumination invariants to deal with illumination changes. Firstly, two-dimensional wavelet transform is performed as a preprocessing step, where low frequency coefficients of a face image are extracted to compress the image, weaken the influence of noises and reduce redundant information. Then, ordinal features are extracted based on an improved operator. The improvements are contributed to by the following two aspects. Firstly, an enlarged functioning domain is employed to increase more information, especially for face recognition tasks with small between-class differences. Secondly, an ellipse-shape operator is proposed to replace the traditional rectangle-shape operator. It is believed that the ellipse-shape operator is more suitable to the local regions of human faces because of the three-dimensional structures and thus, it is more excellent than traditional operator and more relevant information can extracted by using it.The performance of the proposed algorithm is validated by empirical studies conducted with Yale B and ORL face databases, through the comparison with state-of-the-art methods. Face images are grouped according to different illumination conditions. It can be concluded that the proposed method outperforms the compared methods in terms of recognition accuracy and robustness. When single sample image or complex sample images are employed as trainings samples, performances are still stable. Moreover, this algorithm has fewer requirements on training samples and conditions. There are fewer parameters resulting in small time consumption for parameters selection. The extracted ordinal features are able to detailedly describe the regions such as eye, mouth, etc. This algorithm also significantly improves the reliability of feature extraction, especially for large area under shadows.