Author: Xiao, Jun
Title: Age-invariant face recognition based on deep neural network
Degree: M.Sc.
Year: 2018
Subject: Hong Kong Polytechnic University -- Dissertations
Human face recognition (Computer science)
Image analysis -- Data processing
Image processing -- Digital techniques
Department: Department of Electronic and Information Engineering
Pages: 67 pages : color illustrations
Language: English
Abstract: Face-recognition techniques have achieved impressive performances in the past decades, and have been widely used in many applications. In recent years, researchers have directed their attention to face recognition in unrestricted environments. Pose, illumination, occlusion, and aging are four major factors, which degrade the recognition rates of face-recognition algorithms, because they enlarge the intra-person variations. In particular, the aging process has an adverse effect onface recognition, which is unavoidable and irreversible. Aging progression is a highly non-linear function, which results in large intra-person variations by changing local information of facial appearance and becomes worse with time. Existing approaches for age-invariant face recognition can be divided into 2 categories: generative model and discriminant model. The idea of generative models is to compensate for age variation in the model, used for face matching. However, this approach requires a strong assumption. Generally, a generative model has a high computational complexity, which imposes a big limitation on real applications. Compared with generative models, discriminant models show their great power by extracting discriminant features, which are robust to age progression. Usually, discriminant models outperform generative models, and require a smaller amount of computation. The state-of-the-art discriminant models for age-invariant face recognition include Convolutional Neural Network (CNN) based on latent factor analysis. This approach applies the latent factor analysis method to update the fully connected layer of the CNN. Therefore, in this project, we have also investigated and evaluated a discriminant model based on CNN, called "BlockConvNet". In this project, we first evaluate the performance of traditional methods and deep metric learning methods on age-invariant face recognition. First of all, we give a brief introduction to face recognition and the fundamental problems of face recognition: closed-set face recognition and open-set face recognition. We have implemented the open-set face recognition in this project, which is more challenging, compared to closed-set face recognition. We consider some traditional methods, including the PCA, LBP, and SIFT features, and investigate how aging progression affects the recognition accuracy, based on these features, and evaluate their performances for age-invariant face recognition. Based on prior information of facial appearance, "BlockConvNet", a deep-learning structure, is proposed, which combines CNN with metric learning methods, such as center loss for local-feature extraction. We introduce the concept of "age distance", and treat training samples of the young and the old as hard samples. "Block-ConvNet" pays more attention to the hard samples, by adding a penalty into the loss function. This can be viewed as hard-samples mining. Compared with the state-of-the-art method, i.e. latent CNN, our proposed "BlockConvNet" provides an alternative method that directly maps age information to the Euclidean space, and achieves end-to-end learning. From the experiment results, "BlockConvNet" outperforms traditional methods signifcantly and is comparable to the state-of-the-art CNN-based methods.

Files in This Item:
File Description SizeFormat 
991022144625203411.pdfFor PolyU Staff & Students4.04 MBAdobe PDFView/Open


Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: https://theses.lib.polyu.edu.hk/handle/200/9568