16-02-2013, 10:00 AM
THE USE OF MULTI RESOLUTION ACTIVE SHAPE MODELS FOR FACE DETECTION
THE USE OF MULTI RESOLUTION.pdf (Size: 3.03 MB / Downloads: 72)
ABSTRACT
This master thesis examines the use of a multi resolution Active Shape Model (ASM) applied on facial features, utilizing the Viola/Jones face detector.
The method, initially introduced by Cootes, et. al, requires good initial pose parameter values for placing a face model from its local system to the image’s system. This is one of the most critical parts of the process from which the convergence of the method depends on. For this reason, the Viola/Jones detector kicked in, to initially detect the face and subsequently estimate the initial pose parameters for positioning the face model in the search image. The testing of the face detector as well as the quality of the model’s initial position was executed on face images provided by the Milborrow University of Cape Town (MUCT) online database.
For building a face model, a set of training images provided by Cootes was used and the search images were chosen randomly from the same training set.
Experiments made initially on some frontal upright images, showed that the face detector succeeded in all images and the placement of the face model was quite accurate in most cases. Subsequently, the quality of the model fit using the multi resolution active shape model approach, showed that the method converged quite well for the inner part of the face but in the outer part, in some cases, was not that precise.
Introduction
This thesis will describe in detail the implementation of the Active Shape Model method, running in a multi resolution approach, on gray scale images, with the initial pose parameters of the face model estimated by the Viola/Jones face detector. Generally, good mathematical skills are required (especially linear algebra and statistics) and additionally some fundamental knowledge in digital image processing. The reader should not limit himself to the mathematical material presented within this work but use it as a reference for concurring more solid knowledge in the near future.
Motivation
According to the method, when an instance of an object model in a local normalised system is given, initial pose parameter values are requested for placing the model into the image and begin the searching/convergence process. Nevertheless, given some approximate values for translation, rotation and scaling, the process might or might not converge. This means that the method itself is very sensitive and dependent from the given initial values. Having faces as the object; two things are examined within this thesis work:
• Expanding the method to a multi resolution approach. This could improve the efficiency and robustness of the existing algorithm. It starts by searching for face features in a coarse image and when convergence at the current level it’s rescaled and placed as an initial position on the next image level. This leads to a faster algorithm and the probability of failing to detect the correct facial features, is much less.
• Estimating the initial pose parameter values needed for the placement of a face model in the image, using the Viola/Jones algorithm.
Layout of the thesis
In the second chapter, some basics in statistics are introduced. In the third chapter, a very detailed description of the Active Shape Model method is given, firstly explaining the classical approach and then extending it to a multi resolution approach. In the last part of this chapter the Viola/Jones face detector is mentioned. Subsequently in the fourth chapter, the software developed for this thesis is presented. It includes the most important classes and also defines - explains all the functions written for this thesis. In the fifth chapter, results on real face data are demonstrated together with comments about the method used. Finally, the last chapter contains discussions and conclusions with respect to the output results of the method applied and recommendations for improvements.
Background Statistics
This chapter begins by introducing some basics in statistics, with the intention of understanding the Principal Component Analysis approach described later on in this chapter. Principal Component Analysis is a mathematical procedure used in the statistical part of the Active Shape Model method for building (in general) the object model. The reader of this chapter should not restrict himself to the material provided here but consider it as a way of understanding the basic idea behind it and thus be aware how the statistical part of the Active Shape Model works.
Standard Deviation
Standard deviation is a widely used method for measuring the variability in statistics and probability theory. It shows how much variation or “dispersion” the data have from their average (also known as mean or expected value). A low standard deviation indicates that the data points tend to be very close to the mean, whereas high standard deviation shows that the data are spread out over a range of values.
Principal Component Analysis
What is Principal Component Analysis?
Principal Component Analysis (PCA), also known as Hotelling transform is a method that reduces the dimensions of the data by computing the covariance matrix between the data. The first people that started working on this method where Pearson (1901) and Hotelling (1933). Pearson was involved in trying to find lines and planes that best fit a set of points in a n dimensional space. On the other hand, Hotelling tried to increase his “components”, that is the variance in the original variables also known as “principal components”. Both Pearson and Hotelling, came across with the eigenvalue problem (described later on in this chapter), which was hard to solve for an order higher than four and a computer system was needed to process all this information. Today PCA, with the help of powerful computer systems, is a method widely used and established in different fields of applications.
The Basic Principle
As in other transformations (e.g Helmert Transform), PCA tries to transform data from one system to another, where a new set of basis vectors are used. However, in the PCA case, the basis vectors do not remain constant but they vary based on the data being transformed.
Active Shape Models
This chapter describes in detail the Active Shape Model method. It begins with the labeling of the object in training images followed by the alignment of the data sets. It continues with the extraction of some statistical information about the training shapes. Subsequently, a description of the method for getting the gray level appearance information of each model point. Then, the placing of the face model in the testing image to detect the said object. Furthermore, an improved multi resolution approach is described. Last but not least, some basic information of the Viola Jones algorithm is given.
Introduction
Active Shape Models is a method that was developed by T.F. Cootes et.al for detecting known objects in images. Till now, building rigid models of objects for image understanding was well achieved. However, there are cases where objects of the same class are not identical, thus rigid models wouldn’t work, for example the shape of a heart, where it is one object represented from different shapes and sizes. With the method explained in this chapter, new models could be produced from images representing the same object with different shape/ size variations. In addition, Cootes tried to create models that although vary but still preserve the structures of the object class they belong to.
For the method to work, some points are needed that represent the shape of the object within different training examples. These sets are then aligned in order to minimize the variation between equivalent points. From these aligned data sets, a “Point Distribution Model” is created, which gives a mean shape of the aligned shapes and some model parameters that express the different variations within the training set.
Given this model and an image that contains this object, an iterative scheme could be applied that would find the appropriate pose and model parameters which best fit the model in the object.