Face Verification With Statistical Models of Shape and Appearance
CHAPTER ONE
PREAMBLE TO THE STUDY
One goal of the computer vision and machine learning (CVML) group at The Africa University of Science and Technology, Abuja is to build its own code repository from ground up to facilitate research within the group. Many of the algorithms implemented for the repository are not freely available elsewhere or at least not in the organized form implemented in the repository.
Our main goal for this thesis was to build a code repository for the computer vision and machine learning group focusing on research in face recognition and computer animation, and to utilize the code base to perform some experiments in face verification.
CHAPTER TWO
BACKGROUND AND LITERATURE REVIEW
This chapter gives a literature study and background on some of the concepts and the mathematical techniques used in this thesis. We also identified some of the different approaches to face recognition.
Statistical Models
Statistical models are models built from analyzing the appearance of a set of labeled examples of a class of object, such that the model captures the plausible variation in the image structure. A new image can be interpreted by finding the best plausible match of the model to the image data.
The idea is given sufficient training examples; we can capture the variations in a training set made up of instances of the class to be modeled. A statistical model is the model that represents these variations. An example is the shape model in which the set of examples are shapes represented by a number of points called landmarks consistent across all examples.
Statistical Models have the form, where is a vector representing an instance of the object, ̅ is the mean of the set of training examples, is an orthogonal matrix whose columns are unit vectors along the principal modes of variation and is a vector of the model parameters. New examples can be generated which are similar to those in the training set by varying the model parameters (within some constraints).
Examples of statistical models include statistical models of shape which model the variation due to shape, statistical models of texture which model variation due to grey-level intensity of the set of training examples and appearance models which combine the model of shape and texture.
Active Shape Models
Active Shape Models, ASM were proposed by Cootes and his colleagues in a series of papers [6- 10]. In these papers, they proposed methods for building the models and searching with them.
The ASM manipulates the statistical shape model to locate and fit an object of interest in an image.
The ASM has been extensively applied in tracking, and to modeling faces for identification or verification. Baumberg and Hogg [13] used a modified ASM to track people walking and Edwards et. al, [14-16] have applied ASMs for modeling and tracking faces. Other examples of applications of the ASM can be found.
Since the original ASM, there have been many proposed variations. We mention a few here. Rogers and Graham [17] used robust least-square techniques to minimize the residuals between the model shape and the suggested shape instead of the standard least-squares used in basic ASM to address the susceptibility of the standard technique to outliers.
Al-Zubi [18,19] proposed the Active Shape Structural model. It combined ASMs with a model of the structural aspects of the object under consideration.
Van Ginneken et. al. [20] used a k-nearest neighbor classifier instead of the standard ASM profile model search based on Mahalanobis distance.
CHAPTER THREE
ACTIVE SHAPE MODEL
This chapter describes the basic Active shape model, ASM proposed by T.F Cootes [1]. The approach taken is to describe practically the ideas and steps required in building the Shape model and the Active shape model. A more detailed description can be found in the original papers published on this subject [1, 6-10].
Landmarked Points
The statistical model of shape models the variation in the shape of the objects in the training set. In order to represent the shape, the objects need to be labeled by a set of points, called landmarks. The choices of these landmarked points are important and they must be consistently located across all the images. Figure 3.1 below shows examples of two face images with 76 landmark points each.
CHAPTER FOUR
APPEARANCE MODEL
This chapter briefly describes the Appearance model as implemented in this thesis and explains how this is indirectly manipulated by the Active Shape Model to produce parameters that can synthesize a complete image of the object of interest.
The Shape alone does not provide a complete image of an object of interest. In order to have parameters that are sufficient to represent an object, we need to capture both the variation in its shape and its texture, which is the pattern of intensity across the region of the object.
An Appearance model is the model that combines the model of the shape variation and the model of the texture variation in a shape normalized frame. It therefore is able to capture all the information about the object of interest.
CHAPTER FIVE
FACE RECOGNITION
Face recognition for humans is a very trivial task, something that is done most times without conscious thought but for computers it is a very difficult task. Some of the challenges for computers include coping with variations in facial appearance due to facial expression, lighting, pose and age of the person.
Face recognition is a general term that includes both face verification and face identification. Face identification is the task of identifying a face from several already known faces. In a very general way it says, “Give me an image of a face, and I would tell whom it belongs to”. The ability of the system to recognize the face depends on the face being previously known to it. A face identification system should either identify the individual or label the individual as unknown.
CHAPTER SIX
EXPERIMENTS
This chapter describes the experimental setup and then explains in details the experiments conducted and the results obtained.
Database
There are several databases of faces images available [43]; each built under different conditions. For our experiments, we chose the MUCT database because of the number of landmarks and the fact that it is freely available.
REFERENCES
- Cootes, T. F., Taylor, C. J., Cooper, D. H., & Graham, J. (1995). Active shape models- their training and application. Computer vision and image understanding, 61(1),38-59.
- Cootes, T. F., & Taylor, C. J. (2004). Statistical models of appearance forcomputer vision. Imaging Science and Biomedical Engineering, University of Manchester, Manchester M13 9PT, UK March,
- Li, S. Z. (2011). Handbook of face recognition. Springer verlag London
- Viola, P., & Jones, M. J. (2004). Robust real-time face detection. International journal of computer vision, 57(2),137-154.
- Cootes, T. F., Edwards, G. J., & Taylor, C. J. (1998). Active appearance models. In Computer Vision—ECCV‟98 (pp. 484-498). Springer Berlin
- Cootes, F., & Taylor, C. J. (1992). Active shape models—„smart snakes‟. In BMVC92 (pp. 266-275). Springer London.
- Cootes, T. F., Taylor, C. J., Cooper, D. H., & Graham, J. (1992). Training modelsof shape from sets of examples. In BMVC92 (pp. 9-18). Springer
- Cootes, T. F., & Taylor, C. J. (1993, September). Active shape model search using local grey-level models: A quantitative evaluation. In British Machine Vision Conference(Vol. 31, pp. 743-756).
- Cootes, T. F., Taylor, C. J., & Lanitis, A. (1994, September). Active shape models: Evaluation of a multi-resolution method for improving image search. In Proc.British Machine Vision Conference (pp. 327-338).
- Cootes, T. F., Taylor, C. J., Lanitis, A., Cooper, D. H., & Graham, J. (1993, May). Building and using flexible models incorporating grey-level information. In Computer Vision, 1993. Proceedings., Fourth International Conference on (pp. 242-246).