Efficient algorithms for human face modeling

Abstract

The aim of this research is to develop efficient algorithms for modeling face images, which include human face detection, human face tracking, and human face modeling. Human face detection is the first step of all face processing system. In order to achieve a good detection performance level, the system needs to be able to detect the human face accurately and efficiently in a face image. In this thesis, an efficient algorithm for detecting human faces in color images is proposed. The first step in our algorithm is to segment out skin-color regions by using mean-shift algorithm. One of the major challenges to skin-color segmentation is that the human faces may be under poor lighting conditions or under varying lighting over a face region. Our approach considers the distributions of the color components of skin pixels under different illuminations, and the face color regions are identified with the maximum-likelihood technique. Then, the human eyes are detected with color information by using the mean-shift approach. Two eye candidates form a possible face region, which is then verified as a face or not by means of a two-stage procedure with an eigenmask. Finally, the face boundary region of a face candidate is further verified by a probabilistic approach in order to reduce false alarms. Once a face has been detected, it is then tracked in the subsequent video sequence. Facial features are represented using Gabor wavelets, and are then tracked with a modified greedy algorithm. An adaptive template matching method is proposed to adapt the changing appearances during tracking in order to combat the perspective variations of the human face under consideration. The construction of a 3D face model using a similarity measurement is also proposed. Without requiring prior camera calibration, the 3D face model is constructed based on multiple images. An iteration procedure is proposed to estimate the depth of the 3D face model for a human face. Experimental results show that all of these algorithms can detect, track and construct human faces reliably and efficiently.