Face authentication using graph-based low-rank representation of facial local structures for mobile vision applications

Mobile vision systems involve more challenges than ordinary vision systems. In this paper, we propose a novel face authentication approach that considers the difficulties, which exist in mobile vision systems, e.g. lower resolution and quality acquisition, lower storage capabilities, lower computation power, and poor imaging conditions. The proposed authentication approach uses specific fiducial components for authentication purposes. A facial graph model, which involves both of appearance and geometric facial information, is built for face representation. The graph of an image in the gallery set is composed of the low rank matrices of those components as nodes and the mean Euclidean distances between these components as edges. The probe image is represented in a similar way, except that the nodes are represented by the intensity vectors of the components. A weighted matching procedure is performed between the probe and the gallery graphs to make an authentication decision. The proposed system was evaluated using a locally-designed challenging dataset, which is acquired using a cell phone camera. The quality of the captured images of the evaluation dataset is intentionally very low, in order to mimic severe imaging conditions, which a mobile authentication system may encounter. For comparison purposes and as a benchmark test, we evaluate the proposed approach using the FRGC 2.0 dataset. The evaluation results show the potential of the proposed approach in making correct authentication decisions, given very low quality-images. The effectiveness of the proposed approach, in terms of accuracy and memory requirements, versus relevant approaches is demonstrated.