Optimized Face Recognition Using Reinforcement Learning and Deep Learning Feature Extraction
Outlet Title
2025 IEEE 11th International Conference on Big Data Computing Service and Machine Learning Applications (BigDataService)
Document Type
Conference Proceeding
Publication Date
Fall 9-6-2025
Abstract
Face recognition is highly dependent on computer vision, artificial intelligence, and biometrics. Its usage is steadily increasing with systems involving security and user authentication to only cite a few. However, accurate and robust face recognition is not easy to achieve due to varying conditions such as lighting, facial orientations, and obstructions. It also requires substantial computational resources, which is also a constraint for the deployment of face recognition on resource-constrained devices. These limitations show the need to strike a balance between accuracy, robustness, and efficiency when implementing face recognition. This paper proposes a robust and adaptive hybrid feature-extractor-based face recognition method that fuses two lightweight deep learning models: MobileNetV2 and EfficientNetB0. A support vector machine is used for accurate classification. Reinforcement learning, implemented through Q-learning, is used to dynamically optimize the contribution weights α and β for both feature extractors. The system preprocessed the input images, generated hybrid embeddings through a weighted combination of deep features, and changed those weights to achieve optimal performance. The results in both the training and testing data sets were excellent, with an accuracy of 97% and the predictions were verified by confusion matrix analysis with low processing times (27 to 58 ms per step). The main contributions of this work include: the efficient integration of hybrid embeddings for complementary feature representation, a dynamic Reinforcement learning-based weight optimization, and a robustness of the model against variations in lighting, facial orientations, and obstructions. The proposed method is promising for real applications regardless their challenges such as computational overhead and sensitivity to image quality.
Recommended Citation
Harrath, Y., Bhutta, M., Adohinzin, O. and KC, N., 2025, July. Optimized Face Recognition Using Reinforcement Learning and Deep Learning Feature Extraction. In 2025 IEEE 11th International Conference on Big Data Computing Service and Machine Learning Applications (BigDataService) (pp. 218-225). IEEE.