A connexionist approach for robust and precise facial feature detection in complex scenes

We present a technique for robustly and automatically detect a set of user-selected facial features in images, like the eye pupils, the tip of the nose, the mouth centre, etc. Based on a specific architecture of heterogeneous neural layers, the proposed system automatically synthesises simple problem-specific feature extractors and classifiers from a training set of faces with annotated facial features. After training, the facial feature detection system acts like a pipeline of simple filters that treats the raw input face image as a whole and builds global facial feature maps, where facial feature positions can easily be retrieved by a simple search for global maxima. We experimentally show that our method is very robust to lighting and pose variations as well as noise and partial occlusions.