On the Use of Local Fixations and Quality Measures for Deep Face Recognition

Abstract

Automatic face recognition systems have achieved impressive performance in the last decade, thanks to the advances of deep learning techniques. However, these methods are highly dependent on the quality of the acquired face data, which is influenced by several factors. Changes in the face appearance due to variations in illumination, head pose, and visual occlusions, do not affect all regions of the face in the same way. Therefore, a novel approach to extract the quality information from facial regions for building a local deep representation is proposed. In order to determine the local image quality, face images are divided into regions and different quality measures are computed. The good quality regions are used to obtain the final deep face representation. The same strategy is applied for representing face videos by processing every frame of a video sequence as an individual image. Four deep face models are used to evaluate the proposal on five challenging databases, containing both still face images and videos. The performance obtained from the experiments carried out, demonstrates that the proposed approach outperforms the original models on all the tested datasets. The results from the performed longitudinal experimental testing also demonstrates the capability of the proposed deep model to retain robust and highly discriminant textural information from the face image while discarding the data segments corrupted by the considered noise sources.