A Novel Hierarchical Face Recognition Method Based on the Geometrical Face Features and Convolutional Neural Network with a New Layer Arrangement

Abstract

AbstractThis paper represents the face recognition system utilizing a hierarchical arrangement for feature extraction and classification. The proposed method uses the active appearance graph model (AMM) as the first feature extractor and CNN as the second extractor. Also, for each stage, a classifier is considered. The paper investigates the results of two different types of classifiers, SVM and Softmax. The support vector machine shows superior function and outcomes. For the first stage, the AMM extracts six points from the face, then calculates three axillary points based on those to later measure and create five axes or distances on the face as the feature map for each image. It is worth mentioning that the axillary axes are divided by half to eliminate the effect of different facial orientations by the algorithm. In the second stage, for the CNN to give desired results, various forms of data augmentation, such as horizontal flip, shift, scaling, and rotation, are implemented into the algorithm. The color FERET database is used in this research for evaluation. The AMM divides 1208 classes of the FERET into 64 new classes. The biggest new class has 42 members. Applying the algorithm to the FERET database, we compare the accuracy of the system to the other existing methods. Our method shows a better accuracy of 96.35% with Softmax and 97.68% with SVM. Also, compared to the other techniques, the number of required data augmentation is reduced drastically, which translates to lower computational complexity and faster process.KeywordsFace recognitionHierarchicalActive appearance graph modelGeometricalConvolutional neural networkSVMSoftmaxData augmentation