Neural networks for learning human facial features from labeled graph models

A novel method to extract human facial features is proposed based on a two level model-object matching paradigm. A facial feature is modeled by a hybrid technique using labeled graph templates that allows incorporating the relational information of the objects not the models. Gabor filters are used to extract low-level image features to label the vertices of a graph template. A backpropagation artificial neural network is used to extract the optimum template model from examples and to classify image blocks as object or non-object classes. The performance of the algorithm is tested in locating the left and right eyes in a set of human face images with large perspective variations. The attractive feature of the algorithm is the reasonable size of the feature vector in comparison to gray level based methods and the inherent parallelism in both image feature extraction and object detection algorithms.