Eye Localization Based On Haar-LBP Features And FDR-Adaboost Model

Face recognition has become one of the hot spots on biometric identification technologies, and has attracting more and more attention from researchers. Eye localization plays an important role in many automatic face recognition systems, which forms the basis of face image geometric normalization and further face feature detection. Eye localization is performed on detected face image, and it mainly refers to the process of getting eye areas and returning the area centers from given frontal or nearly frontal face images. Therefore the corresponding procedure mainly composes of feature detection, eye area localization and eye centers positioning.Based on the state of art for eye localization, this thesis focused on the research work in such aspects as image feature detection, classification model training, and eye center positioning. The main contributive work of this thesis can be summarized as the following:(l)Eye Feature Detection based on Haar-LBPFor the purpose of effectively description of eye image samples, this thesis proposes a kind of new feature descriptor called Haar-LBP, which makes full advantageous use of Haar and LBP characteristics. Haar-LBP descriptor not only holds such advantages as simple computation and high computational efficiency from Haar features, but also inherits rich object information from LBP. Experimental results show that the performance of eye detection can be improved by means of Haar-LBP features.(2)Classification based on FDR-AdaBoost Training ModelAlthough cascaded AdaBoost is usually applied for image object detection, the AdaBoost strong classifier can be composed of such weak classifiers learned according to minimum false error rate.This thesis proposes a training model based on FDR-AdaBoost, in which Fisher’s Discriminate Ratio(FDR)is used to measure the discriminative power of image features in the course of weak classifier training. By means of FDR-AdaBoost training model, the features with large discriminative power hold the prefer ability for the construction of weak classifiers in those features with minimum false rate. Therefore the final AdaBoost strong classifier is more generative. Furthermore, the multi-value FDR-AdaBoost strong classifier model based on Haar-LBP features has been realized for eye detection, in which an adaptive strategy about weight adjustment is applied to measure the voting of weak learner. Experimental results show the advantageous object detection performance.(3)Eye Center Positioning based on Blended Variance Projection MethodThe final purpose for eye localization is eye center positioning, which also forms the basis for the step of feature detection in face recognition. This thesis proposes a method for eye center position which effectively combining the statistic characteristics of gray value variance projection and gradient variance projection, and therefore get the center position about both eyes.Experiments on famous face database show that the projection method is simple and easy to realize, and holds high accuracy and detection speed, and therefore meet the demands of real-time positioning.