Face Recognition for the Web
Computer Sciences and Information Technology
Introduction
In computer vision, face video recognition remains a very attractive research topic. Face video recognition has been applied in different sectors, some of which include real-time video surveillance as well as in video analytics. Besides, active studies on deep neural networks are currently going on with deep learning models being applied for objection detection, particularly in face recognition. A more highly efficient approach to face recognition that has been applied in recent years is the convolutional neural networks. This discussion explores the various aspects related to face recognition.
Background
Face recognition can be described as the process of verifying or recognizing the identity of a person from a specific video or facial image. In the field of computer vision research, face recognition remains a very important topic that involves numerous applications such as identify recognition, particularly on social media, access controls, and surveillance systems. Signal processing is a field that has benefited immensely from the emergence of deep learning-based techniques. This development has taken place because it can achieve advanced performance in various applications like object classification and face recognition, among others. Today deep learning algorithms have a very advanced performance that has even surpassed that made by humans. Convolutional neural networks (CNN) are a specific type of deep neural network that has led to huge success within the past decade, especially in solving problems involving image classification or object detection. Moreover, CNN is also used in applications of a wide range that includes object recognition and detection, target tracking, among others, such as image recognition.
Recent research findings show that some scholars can apply CNN to recognize video users’ faces (Dogonadze, Obernosterer & Hou, 2020). This development has significantly improved its performance, a factor that makes CNN be an ideal tool for facial classification in different video systems. In recent years, video-based face recognition has been receiving growing attention because of the growing volume of videos that are regularly being captured through various mobile devices and surveillance systems. The current law enforcement requirements and given that videos have rich multi-view information has made is necessary for the development of accurate and robust facial recognition methods for various surveillance systems.
Although the ubiquity that comes with deep learning algorithms has led to the advancement of face recognition technology, especially for the static face images, there is a significant research challenge arising from video-based face recognition techniques. When compared to the static photos taken under very controlled conditions that involve illumination, pose, and expression, individual video frames come with images that are of relatively low quality due to their unconstrained capture environments.
Related Works
Froba and Kublbeck (2019) suggested a technique that utilizes generalized mean to achieve faster feature sets convergence and wavelet transform to help in face recognition during video streams. Froba and Kublbeck (2019) study used a comparative study that involved the analysis of various techniques. This proposed algorithm used different frames that were obtained by tracking face images contained within the video. Identity and feature verifications were undertaken using deep learning architecture. Once this step was complete, the algorithm was tested on PaSC and YouTube databases, which represent the two most popular databases (Froba & Kublbeck, 2019). The findings from these tests demonstrated that learning could be a very effective technique due to its identification accuracy, especially on facial recognition.
Deep learning can be an effective technology that significantly enhances spoofing attack detections. Spoofing attacks can be prevented through spoofing deep learning, which has a huge potential, especially given the fact that is spoofing biometric traits have become very common within the past decade. According to Dogonadze, Obernosterer and Hou (2020), face spoofing attacks should be detected using non-intrusive techniques that involve the use of single frames from replay video attacks that utilize the deep learning technology to boost computer vision. Mixed study approaches have also been used where experimental spoofing attacks detections techniques are applied. The finding from this experiment suggests that the utilization of single frames from replay video attacks that apply deep learning technology recorded better results in face spoofing attack detections in comparison to results from static algorithms. As such, this finding demonstrates that deep learning can be an effective technology that significantly enhances spoofing attack detections.
Qi, Liu and Schuckers (2018) proposed a CNN-based key-frame extraction engine where their system employed the Face Quality, Assessment model. The system tested during the evaluation involved video databases that included ChokePoint, FIA and PaSC. According to the experimental results obtained by Qi, Liu and Schuckers, the KFE engine can be used to reduce data volume while enhancing the FR performance drastically. Moreover, the KFE engine can achieve higher real-time performance when GPU acceleration is employed since it allows HD videos to function properly in real-time application scenarios. This proposed technique has taken place because it can achieve advanced performance in various applications like object classification and face recognition, among others.
Dogonadze, Obernosterer and Hou (2020) proposed a new model that would improve the detection of face forgery through the use of transfer learning arising from face recognition tasks. Additionally, in settings that applied low resolution, it was established that the performance of single-frame detection was quite poor. However, Dogonadze, Obernosterer and Hou tried to use the neighboring frames to enhance middle frame classifications. This evaluation was done on the public Face-Forensics benchmark with the experimental results showing that the model could achieve advanced accuracy.
Şengür et al. (2018) came up with an approach to Face Liveness Detection that utilized Face Liveness Detection applying transfer learning techniques based on CNN’s architectures like AlexNet and VGG16. Şengür’s study explored various deep learning features and compared then to face liveness detection on common ground. The experimental analysis was done on publicly available databases, namely NUAA and CASIA-FASD, which showed that the proposed technique could achieve comparable and satisfactory results.
Yu and Gao (2017) also proposed a new technique referred to as the biometric quality assessment for videos and face images, and they investigated its applicability on FR applications. To achieve this objective, Yu and Gao used light CNN that had the max-feature-map units that made the biometric quality assessment technique more robust to the noisy labels. Yu and Gao’s (2017) study has been being investigated further through experiments done on YouTube and CASIA databases with the results showing the effectiveness of a very high degree to the proposed biometric quality assessment technique.
Fredj, Bouguezzi and Souani (2020) came up with a framework that would allow for the learning of robust face verifications within an unconstrained environment through the use of aggressive data augmentations. The aim of adopting this model was to use large scale data to learn deep face representation with a huge noise and occluded face. Moreover, Fredj, Bouguezzi and Souani (2020) added adaptive fusion of center loss and softmax loss to be used as supervision signals that were crucial in improving performance and conducting final classifications. The results from this experiment demonstrated that the system suggested can achieve comparable performance to other advanced techniques on the YouTube face, and other labeled faces verification tasks.
Finally, Yang, Bulat and Tzimiropoulos (2020) proposed the FAN-Face system, which utilizes features from pre-trained facial landmark-localization networks that assist in the enhancement of face recognition accuracy. In this proposed system, both the features and landmark heatmaps from the pre-trained facial landmark localization network were integrated into the extraction process, applying face recognition features to come up with face-related information while establishing face matching correspondence. Yang, Bulat and Tzimiropoulos (2020) conducted experiments showing how this proposed approach would function when the existing advanced methods were integrated to systematically improve the accuracy of face recognition for a huge variety of the experimental setting available. This development has taken place because it can achieve advanced performance in various applications like object classification and face recognition, among others.
References
Ben Fredj, H., Bouguezzi, S., & Souani, C. (2020). Face recognition in unconstrained environment with CNN. The Visual Computer. doi:10.1007/s00371-020-01794-9
Froba, B., & Kublbeck, C. (2019). Robust face detection at video frame rate based on edge orientation features. Proceedings of Fifth IEEE International Conference on Automatic Face Gesture Recognition. doi:10.1109/afgr.2002.1004177
Lin, M., & Fan, X. (2011). Low resolution face recognition with pose variations using deep belief networks. 2011 4th International Congress on Image and Signal Processing. doi:10.1109/cisp.2011.6100469
Qi, X., Liu, C., & Schuckers, S. (2018). Boosting face in video recognition via CNN based key frame extraction. 2018 International Conference on Biometrics (ICB). doi:10.1109/icb2018.2018.00030
Qi, X., Liu, C., & Schuckers, S. (2018). CNN based key frame extraction for face in video recognition. 2018 IEEE 4th International Conference on Identity, Security, and Behavior Analysis (ISBA). doi:10.1109/isba.2018.8311477
Sengur, A., Akhtar, Z., Akbulut, Y., Ekici, S., & Budak, U. (2018). Deep feature extraction for face Liveness detection. 2018 International Conference on Artificial Intelligence and Data Processing (IDAP). doi:10.1109/idap.2018.8620804
Yang, J., Bulat, A., & Tzimiropoulos, G. (2020). FAN-face: A simple orthogonal improvement to deep face recognition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07), 12621-12628. doi:10.1609/aaai.v34i07.6953
Yu, J., Sun, K., Gao, F., & Zhu, S. (2018). Face biometric quality assessment via light CNN. Pattern Recognition Letters, 107, 25-32. doi:10.1016/j.patrec.2017.07.015
Zhang, F., & Li, Q. (2017). Deep learning-based data forgery detection in automatic generation control. 2017 IEEE Conference on Communications and Network Security (CNS). doi:10.1109/cns.2017.8228705
