Bridging hybrid deep learning detection and lightweight handcrafted features for robust single sample face recognition

Abstract

Single sample face recognition (SSFR) remains a challenging task due to the limitation of having only one reference image per identity, which reduces embedding diversity and decreases robustness under variations of pose, expression, and illumination. This study proposed a hybrid framework that integrates deep learning-based detection through anchor box optimization and non-maximum suppression (NMS) with lightweight handcrafted feature extraction using local binary pattern (LBP). The detection stage leverages deep learning to ensure robust face localisation, while LBP maintains computational efficiency under limited-sample conditions. The training process showed accuracy improvement from 47.5% at the initial epoch to 98.0% at epoch 72, while testing accuracy stabilized at 85-88% with the best value of 87.9%. Evaluation on 48 new facial images achieved 89.6% accuracy, 95.3% precision, 91.1% recall, 93.1% F1-score, and 0.94 area under the receiver operating characteristic curve (AUC ROC). Real-world implementation on Android and iOS-based attendance applications further validated the model, reaching 88.46% accuracy across 52 tests under 50-400 lux illumination. The findings proved that the proposed hybrid design provides improved accuracy and stability compared with previous approaches.