A face recognition framework based on a pool of techniques and differential evolution

• Published in: Information sciences Vol. 543; pp. 219 - 241
• Main Authors: Plichoski, Guilherme Felippe, Chidambaram, Chidambaram, Parpinelli, Rafael Stubs
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 08.01.2021
• Subjects: Continuous optimization; Evolutionary computation; Illumination variation; Natural computing; Pose variation; Single sample per person; Natural computing; Continuous optimization; Single sample per person; Pose variation; Evolutionary computation; Illumination variation
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2020.06.054

•The presentation of an FR framework that aims to identify the best set of preprocessing and feature extraction techniques.•The parameters of selected techniques are also optimized.•The optimization is carried out by the Differential Evolution algorithm.•Illumination and pose variations are approached in two case studies.•The One Sample Per Person (OSPP) real-world condition is approached. Face Recognition (FR) systems are still facing significant challenges when different image issues, such as variation of illumination, pose, expression, and occlusion, are present in captured images. In many situations, it is only possible to obtain One Sample Per Person (OSPP) for training, representing a challenging real-world condition. The proposed FR framework is defined by an optimizer and a pool of preprocessing and feature extraction techniques. The approach makes a pool of techniques available to the optimizer, that can seek the best set of strategies, and tune its parameters. In this work, the FR framework uses the well-known Differential Evolution algorithm as an optimizer, denominated as FR-DE. Two experimental methodologies are employed to assess the performance of the proposed FR-DE framework. First, it is employed a standard dataset separation, and the Yale Extended B dataset is used, which presents severe illumination variation conditions. The second experimental methodology considers the OSPP problem with illumination and poses variations. Also, two well-known datasets are employed, named CMU-PIE and FERET. The proposed approach is compared with some state-of-art algorithms. The comparative analysis suggests that the proposed framework is competitive and suitable for FR systems.