Develop a Hybrid Human Face Recognition System Based on a Dual Deep Neural Network by Interactive Correction Training.

In recent years, the rapid development of artificial intelligence, Internet of Things, and cloud technology has led to the widespread application of face recognition technology, especially in the face recognition system based on deep learning. At present, the face recognition systems only use a single face recognition model for the identity prediction. Furthermore, systems wouldn't perform data training automatically when facial images are captured under monitoring. In the other words, these systems are supposed to manually review the prediction result of face images in order to add training to improve the accuracy rate, because there is no other review method for data training to evaluate the prediction result in these systems. Therefore, this paper developed a method using mutual correction of classification models based on two human face neural network models, called the Dual Face Recognition Model Interactive Correction Training System. Inside the system, we use both FaceNet and OpenFace as the core, to build a module update algorithm. The method is to achieve each set of forecast identity label and confidence for a new face image by both Face Net training models individually. And then the method proceeds to compare with these two sets and mutual correct the results. The new face image with mutual correction identity is used for data training. Therefore, system can execute automatic and dynamic updates, and effectively improve classification accuracy. Moreover, thus system has the advantages of real-time training and limiting model size due to the use of its training process on the classification model. Experimental results show that our system has high performance with a small number of training samples. The system can also automatically improve the accuracy rate and AUC value, and lower error sample rate through the system's modular update algorithm.