Face alignment using a three layer predictor

Face alignment is an important feature for most facial images related algorithms such as expression analysis, face recognition or detection etc. Also, some images lose information due to factors such as occlusion and lighting and it is important to obtain those lost features. This paper proposes an innovative method for automatic face alignment by utilizing deep learning. First, we use second order gaussian derivatives along with RBF-SVM and Adaboost to classify a first layer of landmark points. Next, we use branching based cascaded regression to obtain a second layer of points which is further used as input to a parallel and multi-scale CNN that gives us the complete output. Results showed the algorithm gave excellent results in comparison to state-of-the-art algorithms