TY - GEN
T1 - Develop a Hybrid Human Face Recognition System Based on a Dual Deep Neural Network by Interactive Correction Training
AU - Lee, Pin Xin
AU - Wang, Ding Chau
AU - Tsai, Zhi Jing
AU - Chen, Chao Chun
N1 - Funding Information:
Acknowledgment. This work was supported by Ministry of Science and Technology (MOST) of Taiwan under Grants MOST 109-2221-E-006-199 and 109-2218-E-006-007. This work was financially supported by the “Intelligent Manufacturing Research Center” (iMRC) in NCKU from The Featured Areas Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education in Taiwan.
Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - 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.
AB - 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.
UR - http://www.scopus.com/inward/record.url?scp=85104796275&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85104796275&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-73280-6_47
DO - 10.1007/978-3-030-73280-6_47
M3 - Conference contribution
AN - SCOPUS:85104796275
SN - 9783030732790
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 593
EP - 605
BT - Intelligent Information and Database Systems - 13th Asian Conference, ACIIDS 2021, Proceedings
A2 - Nguyen, Ngoc Thanh
A2 - Chittayasothorn, Suphamit
A2 - Niyato, Dusit
A2 - Trawiński, Bogdan
PB - Springer Science and Business Media Deutschland GmbH
T2 - 13th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2021
Y2 - 7 April 2021 through 10 April 2021
ER -