Real-time face recognition with live detection

Abstract

Human face recognition is one of the most useful techniques for identifying or authenticating a person. Although research on this topic has been conducted for more than twenty years, many problems still remain, and better techniques for facial feature detection and face recognition are needed. Therefore, the objective of this thesis is to devise and develop efficient methods for preprocessing facial images and recognizing human faces. In this thesis, different approaches for facial feature extraction and human face recognition are reviewed. Facial feature extraction is one of the preprocessing steps for automatic human face recognition. Its accuracy will directly affect the performance of the recognition system. In addition, the location of a face, the facial expression and the lighting conditions in an image may be unknown. The head orientation, face scale and the image quality of faces may be different between the query image and the stored image. The recognition procedure will become more difficult and computationally intensive in order to reduce the effect of the above mentioned problems. Therefore, human face recognition is a challenging research topic. In this research, we propose a modified shape model which can adapt to face images under perspective variations. To make the model represent a face more flexibly, the representations of the important facial features, i.e. the eyes, nose and mouth, and the face contour are separated. An energy function is defined that links up these two representations of a human face. In order to represent a face image under different poses, three models are employed to represent the important facial features: the left-viewed, right-viewed, and frontal-viewed models. Furthermore, the genetic algorithm (GA) is applied to search for the best representation of a face image. One of the major difficulties in human face recognition systems is the pose variation problem. Most of the face recognition approaches assume that the pose of an input face is of upright and frontal view. In our work, we estimate the pose angle of the input face image by the shape model parameters, which are derived from a training data set. Then we use Gabor wavelets as local feature information extracted at the facial feature points for classification. The high-dimensional Gabor feature vectors are reduced by the Principal Component Analysis (PCA). The weighting similarity measure based on the pose angle is proposed in classification. The weighting function incorporates class discriminability of feature parameters to emphasize the significance of feature parameters to a particular pose. The face recognition approach proposed in this thesis can provide a reasonable performance level.