Viewpoint invariant model for face detection

摘要

In this work, we have developed a face invariant model for accurate face detection in images where faces can present viewpoint changes. The face model is based on learning a relation between local features and a face invariant. A probabilistic model learned from a training set captures the relationship between the appearance of facial features and the geometry of face invariant. It is then used to infer a face instance in new image. We use the local invariant features which have the high performances to distinguish objects appearance. The facial features are recognised by an appearance classifier which combines an EM classification and a probabilistic matching. Then, the geometrical parameters are predicted to locate face invariants and a hierarchical clustering method corrects the geometric error of the position of invariants. The geometric classification uses an aggregate value to construct clusters of invariants. The probabilities of facial appearance features are computed to select the best cluster and thus to locate face with arbitrary viewpoint in image. We evaluate our generic invariant by testing it in face detection experiments on different databases. The experimental results show that our face invariant model gives highly accurate face localisation.