Method of Feature Vectors Aggregation for Set-Based Face Recognition

The compact representation of face images set provided by feature vectors aggregation for person recognition task is considered. The set-based face recognition increases the reliability of the classification in case of unconstructed image registration environment. Compact representation reduces computational complexity of image sets comparing in feature space. In the paper we propose neural network aggregation method called double quality GhostVLAD, which is based on local cluster aggregation by ghost vector of locally aggregated descriptors and two parameters evaluation by multicolumn network. The double quality GhostVLAD allows to represent the features of all images in the set by intermediate clusters construction and relative importance estimation based on the neural network “quality” evaluation according to classification task. The presented results show that the proposed method improves the accuracy of the person identification by 1% and 4% compared to GhostVLAD and multicolumn network algorithms respectively.