Deep Neural Networks Based Feature Extraction with Multi-class SVM Classifier for Face Recognition
Abstract
Face recognition (FR) is widely applied in biometrics and it is an essential part of a biometric security system that can be run faster than other security methods and can be performed remotely. FR systems have recently achieved encouraging results using deep learning, especially using Convolutional Neural Networks (CNN). FR systems face many challenges in unconstrained environments that reduce their accuracy.
In this study, a deep learning-based feature combination has been proposed for FR to overcome these challenges. The scheme performs feature-level combination by applying two pre-trained CNN models as deep feature extractors, taking into consideration the CNN architectures that have yet achieved the highest results in the ImageNet Challenge. Pre-trained CNN architectures were utilized for an image-based face biometric system by two strategies. In the first strategy, pre-trained GoogLeNet and VggNet models were utilized to extract deep features separately, followed by a multiclass Support Vector Machine (SVM) classifier. In the second strategy, a feature-level combination was used between two feature vectors extracted by pre-trained GoogLeNet and VggNet models followed by a multiclass SVM classifier.
The proposed system is implemented by using MATLAB 2020b, and to evaluate the performance of the proposed approaches, recognition accuracy is used as an evaluation metric. Three experiments were conducted using different datasets, which included several challenges and limitations: VggFace2, LFW, Essex, and ORL datasets which showed the efficacy of the proposed approaches. Additionally, the combination approach between two CNN-based models improves performance. The combination strategy, in particular, yields accuracy in the range of 95.33% to 99.29% on all datasets. The proposed system was compared to existing models in terms of the LFW, and ORL datasets, the findings showed that the proposed system outperformed most current models in terms of accuracy.
