CAN THE COMBINATION OF FACIAL FEATURES ENHANCE THE PERFORMANCE OF FACE RECOGNITION? 10.5455/jjcit.71-1689717889 Djellab Issam,Laimeche Lakhdar,Redjimi Mohamed Classifier combination,Deep learning,Ensemble CNN,Face recognition,Machine learning 243 79 18-Jul.-2023 7-Sep.-2023 and 3-Oct.-2023 3-Oct.-2023 In recent years, researchers have investigated into various approaches of data combination for face recognition, opening up a novel path of exploration aimed at enhancing recognition reliability by capitalizing on the synergy inherent in diverse data sources. This paper implements a comprehensive comparison between two combination methods based on the score-level and feature-level combination, to determine which method highly improves the overall system performance. In the initial method called Fusion-based Classifier Combination (FCC), we introduce a new fusion rule based on score-level combination. This novel model comprises three classifiers; each trained utilizing well-established feature extraction techniques: Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG) and Compact Binary Facial Descriptors (CBFD). Instead of adhering to conventional combination rules, such as majority vote or maximum scores, the derived scores from each classifier are merged and then trained using a Multi-Layer Perceptron (MLP) classifier to reach the final decision. In the subsequent method, named Sequential CNN deep learning-based face recognition (S-CNN), we extract high-level features from multiple image regions considered as sequential data, employing an ensemble of Convolutional Neural Networks (CNNs). In this scheme, the fully connected layers of each CNN-based image region are combined and fed into a Deep Neural Network (DNN) tailored for facial recognition. The experimental results obtained from well-known face datasets, including Labeled Faces in the Wild (LFW), Olivetti Research Laboratory (ORL) and IARPA Janus Benchmark-C (IJB-C) highlight the competitive performance of both the proposed multi-classifier combination model and the S-CNN deep-learning model when compared to state-of-the-art methods.