Dimensionality reduction methods and image-to-class distance for face recognition and human gait recognition

Abstract

Biometrics has been a hot research topic in computer vision society in recent decades owing to its broad applications in commercial and government systems. Face and human gait are two of the most important biometrics which possess huge potential to recognize human unobtrusively. The whole recognition procedure can be divided into three steps: preprocessing, feature extraction/processing, and classification. This thesis focuses on the last two steps: feature extraction/processing and classification. For feature extraction/processing step, three different dimensionality reduction methods are proposed based on holistic features and local patch features respectively and a new distance metric is introduced in the classification step. For the classification step, this thesis introduces an enhanced Image-to-Class distance to compute the distance from one probe image to the set of gallery images from the same person based on the local patch features. We formulate this task as an integer programming problem which incorporates the spatial constraint into each local patch feature by only allowing patches in a spatial neighborhood to be matched. Our proposed Image-to-Class distance is demonstrated to be more effective than the Image-to-Image distance and other existing distance measures in face and human gait recognition. To deal with holistic features in the feature extraction/processing step, dimensionality reduction methods have been broadly used to enhance recognition performance. This thesis proposes the new Trace Ratio based Flexible Discriminant Analysis (TR-FSDA) which relaxes the hard constraint in Semi-supervised Discriminant Analysis (SDA) and adds a regularizer to better cope with data points sampled from non-linear manifold in face recognition.