Learning hierarchical face representation to enhance HCI among medical robots
Introduction
Face recognition has attracted widely attention in recent years, which can be used in various intelligent systems, such as smart phone unlocking, online payment and access control system [1], [2], [3]. Face recognition attempts to detect and track human faces from the captured images. It plays a significant role in biological verification. Compared with iris recognition and fingerprint recognition, face recognition improves the convenience and security of the verification system. Besides, face recognition technology has been the focus of both academic and industry research because of its convenient, user-friendly, and non-mandatory. For medical robot applications, as shown in Fig. 1, by accurately detecting the face attributes of the man, the medical robots can assist the robots to accurately diagnose his potential disease.
Face recognition task recognizes faces by comparing the similarity of pairwise facial feature representations. Thus, it can be regarded as a measuring learning task. In general, face recognition consists of four components: face detection, face alignment, feature extraction and feature matching. Face detection is the basis of face recognition. Classical algorithms of face detection including literatures [4], [5], [6], [7]. Face recognition can be divided into face verification and face identification. Face verification is to determine whether the same image comes from the same person, which is a one-to-one matching task. Face recognition refers to the matching between a given image and a registered face database, which is a one-to-many matching task. Traditional face recognition algorithms focusing on low-level features fail to cope with luminance and face appearance changes. Haar-like feature [8] have been widely used in face recognition, where “integral graph” is used to accelerate feature calculation. Although there are significant features of human faces, different conditions, such as light, posture, occlusion and facial expression, will lead to the problem that small space between classes and large space within classes. In recent years, deep learning algorithms have attracted researchers’ interest, and using deep learning method to solve the problem of face recognition has become a hot topic. Face recognition has achieved great progress under constrained conditions, especially under one-to-one verification, the recognition rate is very high. Face recognition has been widely used in access control system or clock system. However, since the variations of human age, gender, and posture, face recognition cannot achieve satisfactory performance under non-constrained conditions. In addition, the performance of human face recognition will sharply decline without large-scale dataset. Thus, it is of great significance to study face recognition with small samples.
Although existing face recognition algorithms have shown effective performance, they are prone to be attacked by face presentation attacks (face-PAs) [9]. Commonly used face-PAs include printed paper, video reply, 3D synthesis face, and silicone/latex masks. For example, one can leverage a photo of a staff to “cheat” the intelligent access control system, which will lead to serious consequences. Even worse, if one leverage a user’s photo to “attack” online payment system, it will lead to the loss of user funds. How to effectively recognize real human faces from fake samples remains a big problem [10], [11], [12]. In this paper, we propose a hierarchical face representation method. More specifically, an entire face is first separated into several patches including eyes, nose and mouth. And a binary facegrid is generated to indicate the accurate position of the key patches in face image. We design a hierarchical CNN architecture to obtain deep representation of face image. Then we leverage PCA to further reduce the feature dimension. Finally, we leverage SVM for face recognition.
Section snippets
Related work
Geometric features of human face were commonly used in face recognition in previous researches [13], [14], [15]. These algorithms considered the spatial relationship among key patches of human faces, such as eyes, nose and mouth. However, these methods fail to consider shape and texture feature of human faces. Feature extraction methods of human face subspace leveraged eigenvector of covariance matrix of face sample set to model and extract features of human faces. Representative work includes
Our proposed method
In this paper, we propose a hierarchical CNN framework for face recognition by learning deep representation. The pipeline of our used hierarchical CNN architecture is shown in Fig. 2.
Experiment and analysis
In this section, we conduct our experiment on CelebA, LFW [26] and CAS-PEAL [27] datasets. We begin with dataset introduction.
Conclusions
This paper proposes a hierarchical framework for face recognition by learning deep representation. More specifically, we first separate the entire image into several patches including eyes, nose, and mouth, which can exploit key patches for face recognition. A binary facegrid is used to indicate the accurate position of the key patches in face image. The patches are fed into the hierarchical framework to learn the deep representation of the image. Subsequently, we leverage the PCA and SVM
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This study was supported by Key R&D Program of Shandong Province, China (major scientific and technological innovation project) (NO. 2019JZZY011101).
Dianmin Sun was born in Weifang, Shandong, P.R. China, in 1982. He received the MD degree from Tsinghua University, P.R. China. Now, Now, He works in Department of thoracic surgery, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Shandong P.R. China. His research interests include computational intelligence, deep learning and image segmentation algorithms.
E-mail: [email protected]
References (29)
Linear local tangent space alignment and application to face recognition
Neurocomputing
(2007)- et al.
Design of effective neural network ensembles for image classification purposes
Image Vis. Comput.
(2001) - et al.
Face description with local binary patterns: application to face recognition
IEEE Trans. Pattern Anal. Mach. Intell.
(2006) - et al.
Evolutionary pursuit and its application to face recognition
IEEE Trans. Pattern Anal. Mach. Intell.
(2000) - et al.
Neural network-based face detection
IEEE Trans. Pattern Anal. Mach. Intell.
(1998) - et al.
Example-based learning for view-based human face detection
IEEE Trans. Pattern Anal. Mach. Intell.
(1998) - et al.
Face detection, pose estimation, and landmark localization in the wild
Statistical learning of multi-view face detection
- et al.
A general framework for object detection
A face antispoofing database with diverse attacks
Face anti-spoofing based on color texture analysis
Face anti-spoofing based on general image quality assessment
Learning temporal features using LSTM-CNN architecture for face anti-spoofing
Automatic 3D face recognition combining global geometric features with local shape variation information
Cited by (1)
A Deep Learning-Based Robot Analysis Model for Semantic Context Capturing by Using Predictive Models in Public Management
2024, Journal of Global Information Management
Dianmin Sun was born in Weifang, Shandong, P.R. China, in 1982. He received the MD degree from Tsinghua University, P.R. China. Now, Now, He works in Department of thoracic surgery, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Shandong P.R. China. His research interests include computational intelligence, deep learning and image segmentation algorithms.
E-mail: [email protected]
Honghua Zhao was born in Weifang, Shandong, P.R. China, in 1978. He received the doctor’s degree from Beijing Institute of Technology, P.R. China. Now, he works in the School of Mechanical Engineering, University of Jinan, Shandong, P.R. China. His research interest include medical robot, key components of robot.
E-mail: me˙[email protected]
Tao Song was born in June 1983. He is a doctoral supervisor and a young expert of Taishan Scholars in Shandong Province. Deputy Secretary General of the special committee for biological computing and biological information processing, circuit and system branch, Chinese society of electronics, Secretary General of IMCs, deputy editor in chief of cogent engineering, International Journal of adaptive and innovative systems, Editorial Committee, guest editor and reviewer of several SCI journals. He served as vice president of the previous international conferences on Biological Computing (bicta-2012), member of the program committee (bicta 2013), publishing chairman (bicta 2013–2016), Secretary of the previous conferences on membrane computing in Asia (acmc2012), and member of the program committee (acmc 2013–2014). Special report of the 4th acmc 2015 conference and the 1st China membrane computing conference.
E-mail: [email protected]
Aiqin Liu was born in Weifang, Shandong, P.R. China, in 1980. She received the Master degree from Shandong University, P.R. China. Now, she works in School of Information Science and Engineering, Qilu Normal University, Shandong P.R. China. Her research interests include cloud computing, computational intelligence and big data analysis.
E-mail: [email protected]
Jinling Cheng was born in Yantai, Shandong, P.R. China, in 1982. Now, She works in the Digestive Department, Shandong Provincial Western Hospital, Jinan, Shandong, P.R. China. Her research interest include diagnosis of digestive endoscopy, computational intelligence and deep learning.
E-mail: [email protected]
Zhi Liu, Professor, doctoral supervisor. In 2008, he received a Ph.D. in pattern recognition and intelligent system from Shanghai Jiaotong University, a postdoctoral degree from Danish University of science and technology from 2010 to 2012, and a visiting scholar from Columbia University from 2016 to 2017. IEEE member, member of American Optical Society, communication member of pattern recognition Committee of China artificial intelligence society, director of Shandong artificial intelligence society, Shandong “Taishan industry leader”, Shandong “think tank high-end talents”, Jinan “Quancheng special expert”. Won the 7th “Wu Wenjun artificial intelligence science and Technology Award”. Main research fields: image processing, pattern recognition theory and application; multi-source (medical) information fusion; Cyberspace Security (biometric recognition, intelligent intrusion detection, deep mining and traceability of hidden network information).
E-mail: [email protected]
Xin Zhao was born in Gaomi, Shandong, P.R. China, in 1980. He received the Ph.D from Peking Union Medical College, Chinese Academy of Medical Sciences, P.R. China. Now, he works in the Department of Cardiovascular Surgery, Qilu Hospital of Shandong University. His research interest includes the treatment of cardiovascular diseases, the computational intelligence application for cardiovascular surgery, and the big data analysis for cardiac diseases.
E-mail: [email protected]