Abstract
In this thesis, we propose a novel algorithm to automatically extract facial feature points from image. Our approach is base on the statistical models computed from a large number of training images. There are two parts in this thesis, namely face detection and facial feature points extraction. To extract facial features from images, we have to locate the human face first. Our face detection algorithm is based on the multi-templates matching method. The face templates are computed from the collected face images. A hierarchical nearest neighbor network is applied to store the transformed face templates and it can accelerate the matching process between the test image blocks and face templates. The second part in this thesis is the facial feature points extraction, including the corners of eyes and the corners of mouth. We apply the technique of principal component analysis (PCA) to calculate the statistical models for the feature images and facial configurations. We first locate several candidate feature points in each local area on face. Then we consider each combination of these candidate feature points to see whether it is a reasonable facial configuration. Finally, we use the weighted sum of image similarity measure and face shape approximation error to determine the best combination of the candidate feature points.