Face recognition based on deep learning

This research presents the design and implementation of a high-accuracy, real-time face recognition system through a hybrid approach that combines deep learning-based feature extraction with classical machine learning classifiers. The system utilizes a Multi-Task Cascaded Convolutional Neural Network (MTCNN) for face detection, a pre-trained FaceNet (Inception-ResNet-v1) model to generate identity-specific 128-dimensional embedding vectors, and a Support Vector Machine (SVM) for the classification of these vectors. This modular architecture allows for updating only the lightweight and fast-to-train SVM classifier when new individuals are added to the system, thereby eliminating the high computational costs and retraining challenges associated with fully end-to-end deep learning models. Data augmentation techniques such as rotation, brightness variation, and mirroring have been systematically applied to enhance the model's generalization capacity, even with a limited number of original images. Experimental tests have shown that the system achieves a high recognition accuracy of 99,90% on a discrete test set. Furthermore, a learning curve analysis scientifically proves that the model maintains a similar level of performance (99,89%) in cross-validation tests without showing a tendency for overfitting, indicating its high generalization ability. The implemented data augmentation techniques were quantitatively proven to be a fundamental requirement for the system's success, providing a +77,9% increase in the model's Precision metric. Dimensionality reduction analyses (t-SNE, UMAP) have visually confirmed that the feature space generated by FaceNet creates distinct and separable clusters where even a linear SVM can operate effectively. The developed prototype exhibits fluid real-time performance on standard hardware, demonstrating that the proposed hybrid architecture creates a powerful synergy in terms of both high accuracy and practical applicability.