Robust and efficient face recognition via adaptive masking and dictionary learning

Pao Yue-kong Library Electronic Theses Database

Robust and efficient face recognition via adaptive masking and dictionary learning


Author: Feng, Zhizhao
Title: Robust and efficient face recognition via adaptive masking and dictionary learning
Degree: M.Phil.
Year: 2012
Subject: Human face recognition (Computer science)
Machine learning.
Hong Kong Polytechnic University -- Dissertations
Department: Dept. of Computing
Pages: xii, 72 p. : ill. ; 30 cm.
Language: English
InnoPac Record:
Abstract: Biometrics technologies have been more and more widely used in our daily life to identify a person by investigating his/her physiological and behavioral characteristics. Among various biometrics identifiers, face as a distinctive and easy to use biometrics identifier has been widely studied for more than thirty years. However, there are still many challenging problems remaining in face recognition. In this thesis, we propose two coding based robust and efficient face recognition schemes which aim to solve the face occlusion problem and discriminative feature extraction problem, respectively. Face occlusion refers to that the query images are occluded partially by glasses, scarves, or irrelevant images. The occlusion sometimes covers a large part of the frontal face image, which greatly compromises facial feature extraction in conventional methods, leading to failure in face recognition and limiting their applications to practical systems. It is known that occluded pixels usually have high coding errors when representing a face image over the non-occluded training samples. Based on this fact, we propose a novel coding residual map learning scheme for fast and robust face recognition, namely Fast Robust Face Recognition via Coding Residual Map Learning based Adaptive Masking (CRMLAM). A dictionary is learnt to code the training samples, and the distribution of coding residuals can be learnt simultaneously. A residual map can then be obtained to detect the occlusions by adaptive thresholding. Finally the face image can be identified robustly by masking the detected occlusion pixels from face representation. The occluded pixels can be approximately located through this method, and thus the recognition rate can be greatly increased comparing with some state-of-the-art face recognition methods. In face recognition, the computational cost is always an important factor to be considered. Face image usually lies in a high dimension space. Although some prevailing face recognition methods can achieve competing recognition result, the large amount of computational cost greatly reduces their availability. In our proposed method, both the face coding residual and the face coding coefficients are modeled by l₂-norm, and thus the time consumed in face representation and occluded pixel detection is low. By our experiments on benchmark and large scale face databases, the total amount of time cost for recognizing one query image is normally less than one second under the Matlab programming environment, which is very fast compared to state-of-the-art robust face recognition methods. Meanwhile, the face recognition accuracy by our method is very competitive.
The high dimension of face image not only leads to high computational cost, but also prevents the discriminative features from being used for face recognition. In fact, there is much trivial information in the face image which is not desirable in face recognition. Many dimensionality reduction methods have been proposed to solve this problem, while the dictionary learning methods can also be used to reduce the redundant information for a more accurate face representation. Dimensionality reduction and dictionary learning are often considered as two separate steps; in this thesis, we propose a joint learning scheme of dimensionality reduction and dictionary learning, namely Joint Discriminative Dimensionality Reduction and Dictionary Learning (JDDRDL). A face projection matrix and a face representation dictionary are learnt simultaneously by one objective function. By JDDRDL, it is expected that the face features could lie in a more discriminative low dimensional space, where a more representative dictionary can be used to code the face features. Since discriminative information is enhanced in both projection and dictionary learning, the proposed method can better handle the small sample size problem in face recognition. When the number of training sample is insufficient, the recognition rate of many dimensionality-reduction or dictionary-learning based face recognition methods will drop a lot. In comparison, the proposed JDDRDL is still able to achieve satisfying recognition result by exploiting effectively the training information. The major contributions of this thesis are summarized as follows: (1) An efficient and robust face recognition scheme is proposed by learning a dictionary and a coding residual map from the training samples, and coding the query sample over the learnt dictionary with adaptive masking. The proposed method is robust to face occlusion but with a low computational cost; (2) A joint discriminative dimensionality reduction and dictionary learning scheme is developed, which is more robust to the small sample size problem and achieves better face recognition results than state-of-the-art methods.

Files in this item

Files Size Format
b25227269.pdf 1.837Mb PDF
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.


Quick Search


More Information