Gabor feature constrained statistical model for efficient landmark localization and face recognition

Please download to get full document.

View again

of 27
49 views
PDF
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Document Description
Gabor feature constrained statistical model for efficient landmark localization and face recognition
Document Share
Document Tags
Document Transcript
   1 Gabor Feature Constrained Statistical Model for Efficient Landmark Localization and Face Recognition Sanqiang Zhao a,* , Yongsheng Gao a , and Baochang Zhang a,b   a  School of Engineering, Griffith University, Nathan Campus, Brisbane, QLD 4111, Australia  b  School of Automation Science and Electrical Engineering, Beihang University, Beijing 100083, China  Abstract:  Feature extraction and classification using Gabor wavelets have proven to be  successful in computer vision and pattern recognition. Gabor feature based Elastic Bunch Graph Matching (EBGM), which demonstrated excellent performance in the FERET evaluation test, has been considered as one of the best algorithms for face recognition due to its robustness against expression, illumination and pose variations. However, EBGM involves considerable computational complexity in its rigid and deformable matching  process, preventing its use in many real-time applications. This paper presents a new Constrained Profile Model (CPM), in cooperation with Flexible Shape Model (FSM) to  form an efficient localization framework. Through Gabor feature constrained local alignment, the proposed method not only avoids local minima in landmark localization, but also circumvents the exhaustive global optimization. Experiments on CAS-PEAL and  FERET databases demonstrated the effectiveness and efficiency of the proposed method.  Keywords:  Face recognition, Elastic Bunch Graph Matching, Flexible Shape Model, Constrained Profile Model, Gabor wavelet. 1.   Introduction Over the past few decades, the issue of automatic face recognition, where computers are still inferior to humans, has attracted great attention in the research community (Zhao et al., *  Corresponding author. Tel.: +61 7 3735 3652; fax: +61 7 3735 5198.  E-mail address: s.zhao@griffith.edu.au (S. Zhao).     2 2003). This is in large due to its extensive applications in various areas such as law enforcement, identity authentication, video surveillance and human-computer interfaces. Existing technologies for face recognition can be roughly classified into holistic approaches and feature-based approaches. Using information derived from the whole face image, holistic approaches, such as Eigenface (Turk and Pentland, 1991) and Fisherface (Belhumeur et al., 1997), are conceptually simple and easy to implement, but their performance is affected by facial expression, pose and illumination changes in practice. On the other hand, feature-based approaches, such as Elastic Bunch Graph Matching (EBGM) (Lades et al., 1993; Wiskott et al., 1997), Line Edge Map (LEM) (Gao and Leung, 2002) and Directional Corner Point (DCP) (Gao and Qi, 2005), extract local information from salient facial features to distinguish faces. Represented by a set of low dimensional local feature vectors, these methods have the advantage of robustness to environmental variations. The EBGM algorithm showed good identification accuracy in the FERET evaluation (Phillips et al., 2000), demonstrating the successful application of Gabor wavelets for face representation (Shen and Bai, 2006). In EBGM, face geometry is represented by an object-adapted (elastic) topology graph, where each node refers to a predefined fiducial point (or landmark). Local features are modeled by a set of Gabor coefficients (known as a  jet  ) and labeled on each node. The comparison of different individuals is conducted through a graph matching strategy. As matching with each model graph is inefficient for large galleries, Wiskott et al. (1997) developed a stack-like structure, called Face Bunch Graph (FBG), to avoid such a process. A bunch  is a set of jets taken from the same node over a small number of model graphs. To build the representing graph for a test face image, a model graph is first placed at an initial location and then is iteratively matched using Gabor jets to optimize its similarity with FBG, instead of with all the model graphs in the gallery. This matching process consists of two consecutive stages: 1) rigid matching  : a model graph is scaled and shifted inside the test image while keeping its grid   3 rigid, which aims to account for global transformations; and 2) deformable matching  : local distortions due to rotations in depth or expression variations are encoded by deforming individual nodes and evaluated by a graph similarity function within a topological constraint. To ensure accurate landmark localization, the time-consuming Gabor convolution is conducted on all the nodes and iteratively repeated in both stages. Although attractive identification accuracy was reported, the expensive computational cost prevents the use of EBGM in many applications. Many efforts have been made to improve the accuracy and robustness of EBGM. Würtz (1997) presented a system in which several practical procedures were integrated to increase the robustness of elastic graph matching against translations, deformations and changes in  background. Tefas et al. (2001) employed support vector machines to derive optimal coefficients that weigh the local similarity values at the nodes of an elastic graph according to their discriminatory power. Zafeiriou et al. (2007) applied discriminant analysis techniques at multiple  phases of elastic graph matching for face verification. Shin et al. (2007) extended elastic graph matching to a generalized EGM (G-EGM) to obtain enhanced performance on globally misaligned faces. Although attractive identification accuracies were reported, the expensive computational cost makes EBGM not suitable to many real-time applications. However, the efficiency issue of EBGM received less attention in the research community.   Considering that most computationally expensive work of EBGM comes from Gabor convolution, Jiao et al. (2003) utilized E-M algorithm to model the Gabor feature distribution for landmark localization. The method achieved a performance better than Active Shape Model (ASM) (Cootes et al., 1995),  but inferior to EBGM. Choi et al. (2008) presented a simplified version of Gabor wavelets (SGWs) and an efficient (but non-EBGM) algorithm for face recognition. Since the SGW was generated through quantizing its corresponding Gabor wavelet, there was a trade-off between computation time and approximation accuracy. In this paper, we present a new Constrained Profile Model (CPM) to reduce the   4 computation time of face landmark localization and recognition. The landmarks are separated into several groups according to the distribution of salient facial features. Two control points  are selected from each group to constrain other non-control points  in the same group, preventing them from sticking into local minima during optimization. In CPM, each control point is modeled  by a Gabor bunch, while all the non-control points are modeled by Normalized Derivative Profile (NDP) (Cootes et al., 1994). Employing both CPM and Flexible Shape Model (FSM) (Lanitis et al., 1997), we form an efficient localization framework (see Fig. 1), in which landmarks can be localized iteratively via two major steps,  searching   and adjusting  . Based on a rough face detection result, the mean shape model of FSM is placed in a test image as the initial state. In the first searching step, a new suggested movement for each landmark is computed. Gabor jets are utilized to find the movements of control points, while the movements of non-control points are calculated by an efficient searching method based on NDP along the normal of the shape  boundary (Cootes et al., 1994). Because NDP is a unidirectional texture model and has less accurate localization characteristics than Gabor jets, an extra procedure, called Gabor feature constrained local alignment  , is conducted to refine the positions of non-control points in each group. In the second adjusting step, the pose parameters (i.e., scaling, rotation and translation) and shape parameters of FSM are adjusted in order to move each landmark as close as possible to the new suggested position. With the new parameters this process is repeated until the shape difference from two consecutive cycles of iteration is less than a predefined threshold. Different from EBGM, Gabor convolution in the proposed method applies only to a few control points, not to non-control points. Therefore the computational cost is significantly reduced. Finally, face recognition is conducted using Gabor jets of control points as these jets have  been calculated previously in the localization process. Gabor features from other landmarks, such as non-control points and interpolated locations, may be introduced into recognition to yield a higher recognition rate with extra computation. However, our experiments revealed that without   5 further using more landmarks, the proposed method is capable of obtaining comparable recognition accuracy to the standard EBGM algorithm with considerably less computation time. Fig. 1.  Framework of the proposed method. The rest of this paper is organized as follows. Section 2 presents a detailed description of the process of landmark localization covering Flexible Shape Model, Constrained Profile Model, CPM searching and FSM adjusting. Section 3 describes the face representation and recognition. Section 4 reports our comparative experimental results on landmark localization and face recognition. The last section concludes the paper. 2.   Landmark localization The proposed framework incorporates both Flexible Shape Model (FSM) and Constrained Profile Model (CPM) for landmark localization. In this section, we first give a brief introduction of FSM, and then present the algorithm of building CPM. The two major steps in landmark localization, CPM searching and FSM adjusting, are described at last. 2.1.    Flexible Shape Model Flexible Shape Model (FSM) (Lanitis et al., 1997) is generated by a statistical analysis of the positions of the feature points from the training set to represent face shape variation due to differences between individuals as well as changes from environmental conditions. It ensures that Initialization Searching Adjusting Converging?Yes Recognition NoCPM FSM Landmark Localization
Similar documents
View more...
Search Related
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks