Accuracy and robustness evaluation of deep learning algorithms in facial recognition systems

To solve the high cost and low accuracy in facial recognition system, a facial recognition system based on deep learning algorithm is designed in this paper. First, the YOLO model is improved by introducing the EfficientNet to enhance the performance of the facial detection model. Second, a feature extraction model based on the loss function of the improved FaceNet is constructed. In the medium test dataset validation, the proposed facial detection model improved the detection accuracy by an average of 26.30 % compared with the YOLOv3 series models. The LFW dataset validation showed that the model achieved 99.54 % accuracy after 90,000 iterations, which was 1.59 % higher than the average of other models. In the mixed dataset, the proposed facial recognition system improved the accuracy by 4.76 % and 8.64 % compared with the existing mainstream systems, respectively. The system shows strong robustness in diverse scenarios with different skin colors, ages, facial occlusions, and expressions. The designed facial detection method has high detection efficiency, and the feature extraction model has superior recognition results. The system can provide real-time recognition in complex scenes such as facial occlusion, meeting real-time requirements.