Combining SURF descriptor and complex networks for face recognition
João Gilberto de Souza Piotto, Fabrício Martins Lopes · 2016
This paper presents a new approach for facial recognition based on complex networks theory. Initially, the network nodes are defined by the pixels selected from SURF technique. The local 8-neighbors of each pixel are used to define the network edges. Therefore, complex networks measures are extracted from the resulting network and thresholds are applied in order to remove weaker edges and the measures are taken again in an iterative way. The number of iterations performed by the method is defined by parameter n allowing explore the dynamics of the complex network measures in different network scales. The extracted measures stored into a single feature vector and it is applied for face recognition. In order to evaluate the proposed approach the classical datasets: Caltech Face Dataset (CFD), Color FERET Database and Head Pose Image Database (HPI) were adopted. In addition, EigenFace, FisherFace and LBPHFace facial recognition methods were adopted in order to compare the proposed approach with competitors methods. The proposed approach shows better accuracy than competitors methods for CFD and HPI, with 95% and 96.3% of accuracy respectively. Regarding the FERET database, the proposed approach presents accuracy of 95.8%, slightly lower than LBPHFace with 97.2%, presenting accurate results for these classical datasets. These results indicate the suitability and robustness of the proposed approach for facial recognition.