Research On Nose Localization Based On Subclass Discriminant Analysis And Fast Corner Detection

Detection and localization of eyes, eyebrows, nose and other face components from human face images are important research subjects in the field of pattern recognition. They are of great significance to the construction of high performance face recognition system. In addition, they make great sense in the study of face detection, face pose estimation, and facial expression analysis and so on.This thesis focuses on the research of nose localization under different illumination and facial expression conditions. A novel nose localization method based on the Subclass Discriminant Analysis (SDA) is proposed in this thesis. In order to improve the accuracy of nose localization, the problem of nostril localization from a face image is also studied, and a nostril detection approach based on fast corner detection and homomorphic filtering is presented. The main research contents and contributions of this thesis are as follows:Firstly, the method for nose localization based on SDA is studied, which can well resolve the problem of nose localization under varying illumination and facial expression conditions. The nose localization approach proposed in our work consists of two SDA-based nose searching stages. The coarse stage detects nose from the whole face region, while the fine stage improves the nose location around the coarsely located position. The major difference between these two stages is the selection of the negative samples used for the training of SDA classifiers. At the coarse nose localization stage, image patches from the whole face region except those centered at nose tip are randomly selected as negative samples; while at the fine nose localization stage, some nose context patches, i.e. image patches around the nose except those being selected as positive samples are used as negative samples. Our method doesn’t use the prior knowledge about the layout of face components on a face and locates the nose from the whole face image. Experimental results on328AR face images with different expressions and illuminations and on161FERET face images with slight pose variation show that our method can achieve high nose localization rates.Secondly, the problem of nostril localization is studied in detail. The grey values of pixels in a nostril are lower than those of pixels around nostril, based on this fact some nostril candidates are extracted using the fast corner detection algorithm. At the same time, our method uses homomorphic filtering to detect nostril candidates. The final nostrils are obtained by selecting and synthesizing the nostril candidates achieved using above two different methods. Experimental results on AR images and FERET images show that our method can achieve high nostril detection rate under different lighting and expression conditions. Based on the relative position relationship between nostril and nose tip, the resulting nostril are further used to improve the nose location, the nose detection accuracy is increased greatly.