Pose-invariant face localization and face recognition based on unsupervised learning of local features

Abstract

In this dissertation, as the key components for practical face recognition systems, pose-invariant face localization model and face recognition model, and their unsupervised learning algorithms are proposed. For pose-invariant face localization, an unsupervised learning algorithm for a facial constellation model is proposed for localization of a face whose pose, rotated angle, scale, and location are various. The constellation model is learned by obtaining the representative features whose perceptual boundaries are adapted in consideration of localization accuracy and generalization power. Face localization is performed through matching features to the representative features, their weighted voting on the hypothesis space of rotation, scale and translation parameters, their fine adjustment based on optimization of locations of feature points, and determination of the pose of a face using the distance from subspace method. Through experiments, it is shown that the obtained constellation model can accurately localize faces whose rotated angle, size, location and pose are random and unknown. For pose-invariant face recognition, the similarity function of each local feature of frontal faces is learned, and then the local features are selected whose within-class similarity is higher than the unselected local features. The similarity function reflects the risk of misclassification. Using the learned similarity function of each local feature, the average within-class similarity of each local feature is obtained that is useful for face recognition. To learn the non-frontal local features corresponding to the learned frontal local features, the local feature prediction model using generalizable associative memory is proposed. The corresponding non-frontal local features are learned whose predicted frontal features using the generalized associative memory have the smallest average prediction errors. Through the proposed unsupervised learning of local features,...