Boosted LightFace: A hybrid DNN and GBM model for boosted facial recognition

Abstract

Facial recognition technology has seen significant advancements, impacting security, surveillance, and personal identification. Deep neural networks have enhanced accuracy and reliability, with integration into everyday devices further accelerating adoption. Researchers explore combining Deep Neural Networks with Gradient Boosting Machines for improved performance. This paper proposes Boosted LightFace, a hybrid Deep Neural Networks and Gradient Boosting Machines model leveraging robust facial recognition and face detection models. The architecture first integrates predictions from five high-performing DNN models. Their distance metrics and classification outcomes are engineered into a tabular dataset of 6,000 image pairs with 13 features. This dataset is then trained using a highly efficient LightGBM model with a low learning rate of 0.01 and 1000 estimators, incorporating an early stopping mechanism, and employing 10-fold cross-validation to maximize generalization. Recent research identifies FaceNet512d as a robust model, surpassing human recognition on the Labeled Faces in The Wild dataset with 98.4% score. Boosted LightFace achieves 99.1% accuracy, surpassing human recognition by 1.6% and outperforming the best single model in LightFace by 0.7%, underscoring the potential of integrating Deep Neural Networks and Gradient Boosting Machines models in advancing facial recognition technology. Furthermore, Boosted LightFace not only outperforms individual models in terms of accuracy but also surpasses them in precision, recall, F1, and AUC scores, highlighting its comprehensive superiority.