Accurate and real-time facial recognition remains a cornerstone of modern computer vision applications. However, existing systems often suffer from high computational costs, limited scalability, and poor adaptability to large, heterogeneous datasets. This paper presents a lightweight yet robust deep learning-based face recognition framework that leverages embedding-based classification using a custom-designed Deep Neural Network (DNN) architecture. The proposed system integrates Dlib’s ResNet-based face embedding extractor with a bespoke DNN classifier, trained and evaluated on both a custom 68-label dataset and the publicly available Labeled Faces dataset comprising 5,817 identities. The framework encompasses a comprehensive pipeline including face detection using HOG and Haarcascade algorithms, landmark-based alignment using 68 facial points, grayscale preprocessing, and extensive data augmentation to enhance generalization. A 128-dimensional facial encoding is used as input to the DNN, which employs dropout regularization and ReLU activation across fully connected layers to optimize classification performance. Extensive experimentation demonstrates the superiority of the proposed model over traditional and pretrained models. On the 68-label dataset, the DNN achieved 99.37% accuracy, 98% precision, and 89% F1-score, outperforming both Dlib+SVC and conventional methods such as LBPH, Eigenface, and Fisherface. Furthermore, for the large-scale 5,817-label dataset, it attained a 94% accuracy, significantly higher than the 54% achieved by the pretrained SVC model. Real-time testing using a live camera further confirms the framework’s practicality, delivering high-confidence recognition with low latency. This research not only bridges the gap between lightweight deployment and high accuracy but also paves the way for scalable and efficient face recognition in resource-constrained environments.