Up: Publication homepage

Feature-Based Face Recognition Using Mixture-Distance

Ingemar J. Cox - Joumana Ghosn - Peter N. Yianilos1

Abstract:

We consider the problem of feature-based face recognition in the setting where only a single example of each face is available for training. The mixture-distance technique we introduce achieves a recognition rate of $95$% on a database of 685 people in which each face is represented by $30$ measured distances. This is currently the best recorded recognition rate for a feature-based system applied to a database of this size. By comparison, nearest neighbor search using Euclidean distance yields $84$%.

In our work a novel distance function is constructed based on local second order statistics as estimated by modeling the training data as a mixture of normal densities. We report on the results from mixtures of several sizes.

We demonstrate that a flat mixture of mixtures performs as well as the best model and therefore represents an effective solution to the model selection problem. A mixture perspective is also taken for individual Gaussians to choose between first order (variance) and second order (covariance) models. Here an approximation to flat combination is proposed and seen to perform well in practice.

Our results demonstrate that even in the absence of multiple training examples for each class, it is sometimes possible to infer from a statistical model of training data, a significantly improved distance function for use in pattern recognition.

Keywords -- Face Recognition, Mixture Models, Statistical Pattern Recognition, Improved Distance Metrics.


Introduction

Research towards automatic face recognition began in the late $1960's$ and divides roughly into two lines of inquiry: feature based approaches which rely on a feature set small in comparison to the number of image pixels, and direct image methods which involve no intermediate feature extraction stage. There are distinct advantages to both approaches and this is discussed further in Section 2 where previous work is summarized.

This paper's general motivation is to better understand what is limiting the performance of feature based systems. The structure of such systems varies widely but three major components may be identified: the definition of a feature set, the extraction of these features from an image, and the recognition algorithm. We focus on feature sets derived from the location of anatomical features in frontal or nearly frontal views. Our particular feature set definition involves 30 distances derived from 35 measured locations. Our main interest is in the recognition algorithm's effect on performance, so these 35 locations were determined by human operators and recorded in the database. That is, errors associated with feature extraction were kept to a minimum to highlight the errors due to the recognition algorithm, although, in principle, automated feature extraction is possible. This is discussed in greater detail in Section 3 where our experimental database and framework is described.

If many images of each person are available, then each individual may be considered a pattern class, and one can directly apply the methods of statistical pattern recognition to build a model per person. A common approach models each class as a normal density, so for each person there is a corresponding mean feature vector and covariance matrix. The probability of an unknown pattern conditioned on each model is then easily computed. Using a prior distribution on the individuals in the database (flat for example) the classification task is completed in the standard Bayesian fashion by computing the a posteriori probability of each person, conditioned on observation of the query. If the computation is performed using $\log$ probabilities, it is slightly less expensive computationally and the distance metric is the well known Mahalanobis distance.

Given a large number of images for each person this approach would further illuminate the recognition capacity of a given feature set. However in practice we do not always have a large number of images of each individual. In fact, it is not uncommon to have only a single training example for each person, and it is this data sparsity that distinguishes the current work from the traditional class modeling framework. In this setting we assume that the recognition algorithm consists of nearest neighbor search using some distance function between feature vectors. The key questions are then how does the distance function affect recognition rate, and what can be done to find an effective metric?

Our experimental study uses a database of 685 individuals described further in Section 3. Duplicate images are available for $95$ of these and form the queries we use to measure performance. If standard Euclidean distance is used, $84$% of queries are correctly classified. In statistical terms, Euclidean distance may be viewed as corresponding to the assumption that each pattern vector is a class generator with unit covariance and mean coinciding with the pattern. Despite the sparsity of data, it would be surprising indeed if there is nothing one can learn from the training data to improve upon this assumption. To this end, we introduce the use of mixture-distance functions which are obtained by first modeling the training data as a mixture of normal densities, and then using this model in a particular way to measure distance. Our method increases recognition performance to $95$%, the highest recognition rate for a feature-based system applied to a database of this size. These functions are discussed later in Section 4 and explored in greater detail in [21].

The use of mixture-distances immediately presents two model selection problems: selection of the number of elements in the mixture, and the more basic but sometimes ignored problem of choosing between first and second order statistics for each component Gaussian. In both cases, we implemented a very simple flat prior approach which our experiments showed performs as well as the best individual model, as described in Section 5. The results of our experiments are covered in Section 6. Finally, Section 7 consists of concluding remarks and suggested areas for further study.


Previous work

While research in automatic face recognition began in the late $1960's$, progress has been slow. Recently there has been renewed interest in the problem due in part to its numerous security applications ranging from identification of suspects in police databases to identity verification at automatic teller machines. In this section, we briefly describe related work. We coarsely categorize approaches as either feature based, relying on a feature set small in comparison to the number of image pixels, or direct image methods which involve no intermediate feature extraction stage. Of course, direct image methods may also extract features but the distinction is that such features change significantly with variations in illumination. By contrast, the feature based classification is intended to categorize techniques that are robust to illumination conditions.

Direct methods included template matching [1] and the more recent work of Turk and Pentland [17] on ``eigenfaces''. Template matching is only effective when the query and model images have the same scale, orientation and illumination properties. This is a very restricted regime that is unlikely to be found in many operating environments. Although recently Brunelli and Poggio [2] compared a template matching scheme similar to Baron's [1]with a feature-based method on a database of $47$ individuals and found their template matching method to be superior, no generalization can be drawn from these results which are ``clearly specific to our task and to our implementation''.

Turk and Pentland [17] have proposed a method of face recognition based on principal component analysis. Each image of a face maps to a single point in a very high-dimensional space in which each dimension represents the intensity of an image pixel. They then use principal component analysis to find the low-dimensional projection of this space that best represents the data. Using simple nearest neighbor classification in this space Pentland, Moghaddam and Starner [12] report accuracy of 95% on a data base containing about $3000$ different faces. However, all images in this test seem to be taken with little variation in viewpoint and lighting, although with significant variation in facial expression. Since the method is similar to, although more computationally efficient than correlation based on pixel intensities, these results are consistent with Moses et al's [10] conclusions that correlation methods are relatively insensitive to variations in facial expression. Moses has found that correlation methods are much more sensitive to lighting and viewpoint variations, which raises questions about the potential of the eigenfaces approach to extend to these viewing conditions. However, see Pentland, Moghaddam and Starner for one approach to handling view variation.

In principle, feature-based schemes can be made invariant to scale, rotation and/or illumination variations and it is for this reason that we are interested in them. Early work in this area was first reported by Goldstein et al [4] in which a ``face-feature questionnaire'' was manually completed for each face in the database. Human subjects were then asked to identify faces in databases ranging in size from $64$ to $255$ using $22$ features. Interestingly, only $50$% accuracy was obtained.

Subsequent work addressed the problem of automatically extracting facial features. Kanade [7,9] described a system which automatically extracted a set of facial features, computed a $16$-dimensional feature vector based on ratios of distances (and areas) between facial features, and compared two faces based on a sum of distances. On a database of $20$ faces, Kanade achieved a recognition rate of between $45-75$% using automatically extracted facial features. It is interesting to note that when our mixture contains just a single Gaussian, and only first order statistics are employed (the off-diagonal covariance entries are ignored), our approach reduces to Kanade's early work using Euclidean distance weighted inversely by the variance of each feature.

Perhaps because it was perceived as difficult to automatically extract $2$-dimensional facial features, significant effort has been directed towards using face profiles [5,6,8]. In this case, the automatic extraction of features is a somewhat simpler one-dimensional problem. Kaufman and Breeding reported a recognition rate of $90$% using facial profiles, but this was on a database of only $10$ individuals. Harmon et al reported a recognition rate of $84$% on a $121$ individual database using a Euclidean distance metric. Recognition rates of almost $100$% are claimed using a classification scheme based on set partitioning and Euclidean distance. However, these experiments did not maintain disjoint training and test sets. Subsequent work by Harmon et al [5] did maintain a separate test set and reported recognition accuracies of $96$% on a database of $112$ individuals.

Kaufman and Breeding [8] compared their results with human recognition of facial profiles and found that human performance was not significantly better. This comparison highlights an obvious problem: what is the classification capability of a set of features? This is clearly a fundamental question, especially since it is unclear what features the human visual system uses. After all, no amount of subsequent processing can compensate for a feature set that lacks discrimination ability. Perhaps because of this, most previous work has concentrated on investigating alternative face representations while paying little attention to the subsequent recognition algorithm. In fact, the role of the recognition algorithm has not been adequately addressed in the face recognition literature, especially for moderately large databases $(>100)$. In this paper we begin to do so by examining the recognition rate of a $30$-dimensional feature vector on a database of $685$ faces.


Experimental Database

Figure 1: Manually identified facial features.
\begin{figure}\centering  \psfig{figure=faceFig.eps,height=2.50in,width=2.0in} \end{figure}

Figure (1) shows the $35$ points that were manually extracted from each face and Table (1) lists the 30-dimensional feature vector computed from these facial features. We followed the point measurement system of [19] since the Japanese portion of our database consisted of measured feature values only, i.e. the original intensity images were unavailable. All distances are normalized by the inter-iris distance to provide similarity invariance.

Our model database of $685$ images is an amalgam of images selected from several different sources as described below2: 1) $20$ images from the UCSB database created by B.S. Manjunath of UCSB, 2) $24$ images from Weizmann Institute database which was obtained from public domain ftp site from Weizmann Institute, courtesy of Yael Moses, 3) $12$ images from the MIT database which was down-loaded from the public ftp site at MIT, 4) $533$ images from the NEC database obtained from NEC, Japan, 5) $16$ images from the database provided by Sandy Pentland of MIT Media Lab, 6) $80$ images from the Feret Database, courtesy of the Army Research Laboratory. The query database consists of $95$ images from the following sources: 1) $18$ images from UCSB database, 2) $23$ images from the Weizmann Institute database, 3) $11$ images from the MIT database, 4) $43$ images from the NEC, Japan database. Each element of the query database represents a second frontal view of someone in the model database. Its size was severely limited by the availability of such images.


Table 1: The $30$-dimensional feature vector.
Feature Distance
1 0.5 * ( (1,2) + (11,12) )
2 0.5 * ( (5,6) + (15,16) )
3 (3,13)
4 (24,25)
5 (29,30)
6 (34,35)
7 (26,34)
8 (28,35)
9 (26,28)
10 (27,31)
11 (27,32)
12 (32,33)
13 (23,31)
14 (21,22)
15 0.5 * ( (13,25) + (3,24) )
16 0.5 * ( (25,30) + (24,29) )
17 0.5 * ( (30,34) + (29,35) )
18 0.5 * ( (1,22) + (11,21) )
19 (10,19)
20 0.5 * ( (2,9) + (12,20) )
21 0.5 * ( (9,10) + (19,20) )
22 0.5 * ( (11,19) + (1,10) )
23 0.5 * ( (6,7) + (16,17) )
24 0.5 * ( (7,8) + (17,18) )
25 0.5 * ( (18,19) + (8,10) )
26 0.5 * ( (18,20) + (8,9) )
27 (11,23)
28 (1,23)
29 0.5 * ( (1,28) + (11,26) )
30 0.5 * ( (12,13)+(2,3) )



Mixture Distance Functions

Given a database of facial feature vectors $Y=\{y_i\}$, each corresponding to a different person, and a query $q$ consisting of a facial feature vector for some unidentified person assumed to be represented in $Y$, our objective is to locate the $y_i$ corresponding to $q$. In the absence of error, and assuming no two people are exactly alike, we would have only to search $Y$ for an exact match to $q$. But in practice $q$ will not match anything in $Y$ perfectly because of many sources of error. These include feature extraction errors associated with the human or algorithm which constructed the feature vector from a photograph, variation in the subject's pose, unknown camera optical characteristics, and physical variation in the subject itself (e.g. expression, aging, sickness, grooming, etc.) Clearly the nature of these error processes should influence the way in which we compare queries and database elements. The difficulty lies in the fact that we can't directly observe them given that only a single example of each person exists in $Y$.

In this section we begin with a formal discussion but at a general level in order to establish a clear conceptual framework. Certain simplifying assumptions are then made which lead to a practical approach to the problem of inferring something about the error processes at work in our data. The final result is then a simple formula for comparing queries with database elements in the presence of error.

We imagine the observed feature vectors, whether database elements or queries, to be the result of a two-stage generative process. The first stage ${\cal P}$ generates platonic vectors $p$ which are thought of as idealized representations of each pattern class - in our case the facial features of distinct humans. The second stage is the observation process which generates the vectors we ultimately observe. The first stage corresponds to inter-class variation, i.e. between people, while the second stage captures intra-class variation. The nature of the second process depends on $p$ and we therefore denote it as ${\cal O}_p$. We will further assume that each ${\cal O}_p$ is a zero mean process, which conceptually, adds observation noise to the platonic vector at its center.

The probability $\Pr(q\vert p)$ that a query $q$ was generated by a particular platonic $p$, is then computed by forming the vector difference $q-p$ and evaluating ${\cal O}_p$. This suggests the notation: $\Pr(q\vert p) \mbox{\msam ,}{\cal O}_p(q-p)$. Similarly the probability $\Pr(y_i\vert p)$ that a particular database element $y_i$ was generated by $p$ is ${\cal O}_p(y_i-p)$. Finally the probability $\Pr(p)$ of $p$ itself is just ${\cal P}(p)$. To judge how similar $q$ and $y_i$, the approach taken in [21] is to focus on the probability of the 3-way joint event consisting of the generation of $p$, followed by its observation as $q$, followed by a second independent observation as $y_i$. Integrating over $p$ then gives the probability that $q$ and $y_i$ are independent observations of a single platonic form.

Our first simplifying assumption is that the $y_i$ are considered to be platonic. This eliminates the integral above and is actually the assumption implicit in most nearest neighbor pattern recognition methods. It amounts to imagining that the query is a an observation of the database element - not of some third (mutual) platonic element. One hopes that $y_i$ is not too far from its $p$, and that the distribution of observations about $y_i$ therefore approximates the distribution about $p$. So now we focus on the matter of attributing to $y_i$, an observation process ${\cal O}_i$. Having done this we can then compute $\Pr(q\vert y_i)$ for each $y_i$. We will classify $q$ by choosing the largest such probability. This is easily seen to be the same as maximizing $\Pr(y_i\vert q)$ with a flat prior on $Y$.

The mixture-distance method we introduce may be viewed as a particular way, but by no means the only way to arrive at an ${\cal O}_i$ for each $y_i$. Perhaps the simplest such assignment gives each $y_i$ an identical ${\cal O}_i$ consisting of a zero mean Gaussian process with unit covariance. Computing probabilities as logarithms reveals that this is exactly the same as the use of ordinary Euclidean distance and performing a nearest neighbor search. It is helpful to visualize these ${\cal O}_i$ as hyper-spheres of identical dimension corresponding to the unit distance (or equi-probability) surface arising from the process. In this simple case such a sphere is located about every database element so that the nature of the distance function employed is the same everywhere in space - and in every dimension.

In contrast to this simple assignment, we may also consider the ideal case in which each $y_i$ is associated with its true error process. Assuming these processes are zero mean Gaussian, then the ${\cal O}_i$ may now be visualized as a hyper-ellipsoids of various sizes and shapes surrounding the $y_i$. Unfortunately, as observed earlier, we don't have enough data to reconstruct this picture and must therefore turn to techniques which infer something about it from the few data points available. The mixture-distance technique is such a method which makes its inference based only on the distribution of $Y$.

Suppose the observation process is extremely noisy - so much so that most of the variation seen in $Y$ is due to noise not to actual differences in facial characteristics. In this extreme case, assuming for simplicity that the noise process is Gaussian, the sample covariance matrix of $Y$ captures mainly the characteristics of the observation process. At the other extreme, if little noise is present, then most of the variation in $Y$ is due to actual differences between individuals. Here there is no reason to expect the sample covariance of $Y$ to tell us anything about the observation process.

The main conceptual argument behind mixture-distance is that if $Y$ is decomposed into a mixture, where each component is thought of as covering some region of space, then within each region, observation noise becomes the dominant component of the empirical distribution. So as the number of mixture components increases, one hopes that that the statistics of each region capture an increasingly accurate estimate of the observation process. The ${\cal O}_i$ assigned to each $y_i$ is then determined by the region into which $y_i$ falls.

Consider the data-rich limit in which $Y$ contains many observations of each person, and the mixture contains as many components as there are distinct individual's. Here, given a perfect unsupervised mixture density estimation procedure, one would express $Y$ as a mixture of densities where each component corresponds exactly to the error process for a particular individual, and is centered at the mean feature value for that individual. In this extreme case, attributing to $y_i$ the error process from its region of space, is exactly the right thing to do. In practice one employs mixtures with far fewer components and hopes that the resulting decomposition of space makes the observation process dominant or at least significant. Said another way, one hopes that before reaching the data-rich case above, the decompositions arrived at by unsupervised mixture density estimation, begin to reveal useful information about the observation processes at work in each region of space.

So returning to our imagined hyper-ellipsoids surrounding each $y_i$, mixture-distance may be thought of as assigning ${\cal O}_i$ based on the mixture component which contains $y_i$. A simplified picture would show the $y_i$ in particular region space surrounded by hyper-ellipsoids selected for that region. The imagery above is a simplification of the true situation because each $y_i$ belongs stochastically, not deterministically, to a given region. The ${\cal O}_i$ assigned to it is then a mixture not a single Gaussian.

Also, in the examples above, we assumed that the actual observation processes were zero mean Gaussian. We remark that given even a single face and multiple observations arising from different feature extraction agents (whether human operators or algorithms), a less restrictive assumption is that the error process is itself a mixture of zero mean Gaussians - one for each agent. We make this remark because it is entirely possible that some of the components identified by unsupervised mixture density estimation, may correspond to different feature extractors not to different kinds of faces. In general the structure discovered might sometimes correspond to semantic labels such as as gender, age, racial identity - but there is no reason to believe that such a correspondence is necessary in order for the hidden structure to lead to an improved distance function.

We now proceed to more formally derive our formula for mixture-distance. A finite mixture model $M$ is a collection of probability models $M_1,\ldots,M_n$ and non-negative mixing parameters $c_1,\ldots,c_n$ summing to unity, such that: $
M(x) = \sum_{k=1}^n c_k \cdot M_k(x)
$.

Let $N_{\Sigma,\mu}(x)$ denote the multi-variate normal density (Gaussian) having covariance $\Sigma$ and expectation $\mu$. When the elements $M_i$ of $M$ are Gaussian, $M$ is said to be a Gaussian or Normal mixture. Given a finite set of vectors $x_1,\ldots,x_m$, the task of estimating the parameters of a normal mixture model which explains the data well, has been heavily studied. The well known expectation maximization method (EM) [13] is perhaps that best known approach and we adopt it for our experiments using k-means clustering to provide a starting point.

We now assume that an $n$-element normal mixture model $M$ has been built to model the database elements $\{y_i\}$ and we refer to this as the empirical distribution. Each mixture element $M_k$ is a normal density $N_{\Sigma_k,\mu_k}$ and we note by $\bar{M_k}$ the zero mean density $N_{\Sigma_k,0}$. So $\Pr(x\vert M_k) = \Pr(x-\mu_k\vert\bar{M_k})$. The system's query to be classified is denoted $q$. Using the mixing probabilities $\Pr(M_k)$ obtained from EM, we may then compute the a posteriori component probabilities $\Pr(M_k\vert x)$. These may be thought of as a stochastic indication of $x$'s membership in each of the mixture's components. We will attribute to each $y_i$ an ${\cal O}_i$ which is a mixture of the $\bar{M_k}$ determined by these stochastic membership values. This is explained best by the derivation which follows:


$\displaystyle \Pr(q\vert y,M)$ $\textstyle =$ $\displaystyle \frac{1}{\Pr(y\vert M)} \cdot \Pr(q \cdot y \vert M)$ (1)
  $\textstyle =$ $\displaystyle \frac{1}{\Pr(y\vert M)} \cdot \sum_{k=1}^n
\Pr(q \cdot y \vert M_k) \cdot \Pr(M_k)$  



  $\textstyle =$ $\displaystyle \frac{1}{\Pr(y\vert M)} \cdot \sum_{k=1}^n
\Pr(q\vert y,M_k) \cdot \Pr(y\vert M_k) \cdot \Pr(M_k)$  
  $\textstyle =$ $\displaystyle \frac{1}{\Pr(y\vert M)} \cdot \sum_{k=1}^n
\Pr(q-y\vert\bar{M_k}) \cdot \Pr(y\vert M_k) \cdot \Pr(M_k)$  
  $\textstyle =$ $\displaystyle \sum_{k=1}^n \Pr(q-y\vert\bar{M_k}) \Pr(M_k\vert y)$  

where $\Pr(q\vert y,M_k) = \Pr(q-y\vert\bar{M_k})$ and


\begin{displaymath}
\Pr(M_k\vert y) = \frac{\Pr(y\vert M_k) \Pr(M_k)}{\sum_{i=1}^n \Pr(y\vert M_i) \Pr(M_i)}
\end{displaymath}

It is this formulation we use for all of our experiments. In [21] various more complicated expressions are given corresponding to weaker or different assumptions. Finally we observe that in the case of one mixture element $n=1$, mixture-distance reduces to the Mahalanobis distance from the query $q$ to average face $\mu$.

Efficient Computation

Observe first that the term $\Pr(M_k\vert y)$ does not depend on the query and may therefore be pre-computed and recorded as part of the database. Next recall from the basic theory of multi-variate normal densities and quadratic forms, that for each mixture element $M_k$ we may find a basis in which the density's covariance matrix is diagonal. This is of course accomplished via unitary matrix $E_k$ whose rows consists of the eigenvectors of $\Sigma_k$. If the vectors $E_k y_i$ are all recorded in the database as well, and $E_k q$ is computed before the database search begins, then the computation time for mixture distance becomes linear, not quadratic in the dimension of feature space. Note, however, that $k$ vectors must be stored for each database element $y_i$. This storage requirement can be reduced by the ``hard VQ'' approximation.

The Hard VQ Approximation

Focusing again on $\Pr(M_k\vert y)$ we may make another simplifying assumption in order to further reduce computation and storage space. Note that $\sum_{k=1}^n \Pr(M_k\vert y) = 1$. The assumption which is typically referred to as Hard VQ (where VQ stands for ``vector quantization''), consists of replacing this discrete probability function on $\{M_k\}$ by a simpler function which assumes value $1$ at a single point where the original function is maximized, and zero elsewhere. Conceptually this corresponds to hard decision boundaries and we would expect it to affect the resulting computation significantly only when a point $y_i$ is near to a decision boundary. We will then refer to the original formulation as Soft VQ.

Space savings result from the hard VQ assumption since we must now record in the database, each $y_i$ expressed in only a single Eigenbasis corresponding to the distinguished $M_k$ (which must also be identified). This scheme is then a linear time and space prescription for mixture-distance computation.


Model Selection Techniques

In any statistical modeling approach in which the size and nature of the models (their configuration) used may vary, one faces the model selection problem.3 The objective of model selection is, of course, not to better model the training data, but to ensure that the learned model will generalize to unseen patterns. Many approaches to this fundamental problem have been described in the literature. The key objective is always to prevent the model from becoming too complex, until enough data has been seen to justify it. The simplest result one might hope for is to somehow select a single configuration which is appropriate for the data and problem at hand. Another solution consists of finding a probability function (or density) on configuration space. Here ``selection'' becomes a soft process in which one merely re-weights all of the possibilities. The first objective is clearly a special case of the second. In this paper we will ultimately adopt a very simple selection strategy, but, as motivation, first discuss the problem at a more general level.

Another subtlety associated with the term ``selection'' is that it seems to imply that the final result is at most as good as the best individual model. This is not the case. A blended model can be better than any of its constituents as shown in the experimental results of Section 6. A simple example serves to illustrate this point. Suppose a timid weather-man $A$ predicts rain and sun each day with equal probability while a second sure-of-himself weather-man $B$ always issues certain predictions, i.e. $100$% of rain or $100$% chance of sun. Further assume that $B$ is correct $2/3$ of the time. If the objective is to maximize the probability assigned to long a series of trials, then it is easily verified that one does best by blending their predictions, placing weight $2/3$ on $A$ and $1/3$ on $B$.

For simplicity we will assume a discrete setting in which selection is from among $M^1,\ldots,M^m$. We seek non-negative values $d_1,\ldots,d_m$ summing to unity, such that $M = \sum_{\ell=1}^m d_\ell M^l$ represents a good choice. Here $M$ is of course our final model and each $M^\ell$ may themselves be complex models such as Gaussian mixtures. One approach to model selection consists of Bayesian update in which the starting point is a prior probability function on the configuration patterns, and after each training example is predicted by the model, a posterior distribution is computed.

Our purpose in this paper is to explore the basic effectiveness of the mixture-distance approach so we have confined ourselves a very simple form of model selection which amounts to simply using a flat initial prior and not bothering to update it. That is, we assume all configurations have equal probability and mix them (average) accordingly. Bayesian learning and other approaches may be evaluated in future work.

Selecting between first and second order models

A first order Gaussian model $M_1$ has a diagonal covariance matrix containing estimates of the individual feature variances and a mean vector consisting of an estimate of the distribution's expectation. The second order model $M_2$ matches $M_1$ except that its off-diagonal covariance entries may be non-zero and represent estimates of the feature's second order statistics. The second order model has of course many more parameters, so if limited data is available one should worry about its ability to generalize. Moreover, when forming Gaussian mixtures, the available data are essentially parceled out to mixture elements - further exacerbating the problem.

A mixture is just $M = (1-f) \cdot M_1 + f \cdot M_2, f \in
[0,1]$. Consider this mixture generatively where one first chooses $M_1$ with probability $1-f$ or $M_2$ with probability $f$, and then draws a sample at random according to the selected model. The covariance matrix of the resulting data vectors is easily seen to be $\Sigma_1 + f \cdot (\Sigma_2 -
\Sigma_1)$. This is just $\Sigma_2$ with it's off diagonal elements multiplied by $f$.

Now unfortunately $M$ is not necessarily Normally distributed despite the nature of $M_1$ and $M_2$, so we employ the known statistical expedient in which one approximates a mixture with a single Gaussian. This leads to the following heuristic which we employ in the experiments to follow: The off-diagonal elements of the sample covariance matrix are multiplied by $f$ to form a single Gaussian model which approximately selects between first and second order statistics.

This is of course exactly the ML estimate for the parameters of a single Gaussian trained from the mixed distribution. The introduction of the $f$ parameter has another practical benefit. It we require $f \in [0,1)$ then the resulting covariance matrix cannot be singular unless some feature has exactly zero variance. Our experiments will focus on three natural values: $f=0$, $f=\frac{1}{2}$ and $f \approx
1$.4 We also point out that $f$ is employed everywhere in the statistical process including the maximization step of EM.

Selecting the number of mixture components

It is difficult to know a priori how many mixture elements should be used to describe our database of facial features. Again there are many approaches to this problem but we adopt in some sense the simplest by fixing only the upper end of a range, and mixing all with equal probability.


Experimental Results

To provide a baseline recognition rate to compare our results to, we applied to simple Euclidean distance metric to the database and obtained an $84$% recognition level.

For simplicity we adopt a flat selection policy $f=\frac{1}{2}$ to decide between first and second order models, i.e. between just the diagonal variance and the full covariance matrices. Table 2 illustrates how $f$ can significantly affect recognition accuracy.

Table 2: The $f$ parameter which selects first vs. second order models has a potent effect on recognizer accuracy.
Mixture Recog. Rate Recog. Rate Recog. Rate
Elements $f=0$ $f=0.5$ $f \approx
1$
1 93% 91% 84%
2 86% 94% 84%
5 88% 94% 81%
mixture of NA 95% NA
1-10 mixtures


Notice that when the mixture consists of a single Gaussian, a first order variance model ($f=0$) is best and the full second order covariance model ($f \approx
1$) is considerably worse. However, for mixture sizes $2$ and $5$, an off diagonal weighting, $f=\frac{1}{2}$, is best while the full second order model with $f =
\approx 1$ still a distant third. The point of Table 2 is not that the flat selection model can increase the recognition rate from $93$% to $94$% but rather that the recognition rate is consistently good using a flat selection, i.e. the recognition rate is both high and the variation across mixture models is low. A flat prior is therefore robust and eliminates the uncertainty associated with any choice of first order variance or second order covariance models.

Figure 2: Recognition accuracy varies considerably with mixture complexity. Both ``soft'' and ``hard'' VQ versions of mixture distance are presented.
\begin{figure}\centering  \psfig{figure=fig_685_95.eps,height=2.75in,width=3.25in} \end{figure}

When using a Gaussian mixture model, the number of mixtures present is often unknown. This is a significant problem since the complexity of the Gaussian mixture also affects recognizer performance in a significant and non-monotonic way. This is illustrated by the graphs of Figure 2 in which models containing $1$ through $10$ Gaussians were tested. The right graph shows the results using ``soft-VQ'' mixture distance, and the left graph corresponds to ``hard-VQ''.

Discussing the soft vector quantization method first, we notice that the peak recognition rate is $94\%$ but that the rate varies considerably as the number of mixture elements changes. Some of this variation might have been reduced if multiple EM starting points were used and the recognition results averaged. However, as in the case of the $f$ parameter above, our experiments highlight the difficulty of model selection. To alleviate this, we again propose a flat stochastic selection scheme, i.e. we assume that each model in the complexity range $1-10$ is equally likely and form a mixture of mixtures. The result is that $95$% accuracy is achieved and this exceeds the performance of any individual model. Once more though, the significance of this results is not just the improvement in recognition rate but also the fact that the best recognition rate is achieved while simultaneously removing the uncertainty associated with mixture selection.

The Hard VQ version of mixture-distance is somewhat attractive if computational cost is an important issue, as described in Section 4. The left graph of Figure 2 shows its performance which, like the soft VQ method, is highly variable with mixture complexity. The best performance $95$% is attained for $5$ mixture elements and exceeds the $94$% maximum level of Figure 2. However when a flat mixture of mixtures was formed as for the soft strategy, performance of $94$% resulted. Again, the conclusion to be drawn is that mixtures of mixtures remove the uncertainty due to variability of recognition rate with mixture complexity while simultaneously providing excellent performance.

Finally we report that limited experiments on the effect of increasing database size suggest that performance declines significantly when only a single mixture element is used, and is far more stable given larger mixtures.


Concluding Remarks

We have demonstrated that the use of a simple form of mixture-distance, along with a simple solution to the model selection problem, increase performance on our face recognition problem from $84$% using Euclidean distance to $95$%. This provides strong motivation for careful consideration when choosing an appropriate metric. A less impressive but still significant increase from $93$% to $95$% was observed when we compare the results of a single first order Gaussian model, with the results using large mixtures of mixtures. Just as importantly, the recognition rate is consistently good using a mixture of mixtures and flat priors on both the order and model selection. In contrast, it was observed that specific selection of a mixture model and order statistics can lead to considerable variations in the recognition rate. The mixture of mixtures is a robust technique that eliminates this uncertainty. Nevertheless, further experiments in the face recognition domain and others will be necessary to evaluate the significance of the contribution made by generalizing to second order models and mixtures.

Given the small size of our query database, and our limited problem domain, it is not possible to conclusively demonstrate the general effectiveness of the mixture-distance approach. Nevertheless, our results suggest that $(1)$ it does lead to significant improvements over simple Euclidean distance, $(2)$ that flat stochastic selection is an effective solution to both model selection problems, $(3)$ that flat stochastic selection significantly reduces the otherwise serious variability of recognition rate with model parameters and $(4)$ that the hard-VQ algorithm compares well with the computationally more expensive soft-VQ.

It is also important to realize that the techniques of this paper are quite independent of the particular feature set we chose for experimentation. In fact, mixture-distances can be applied to more direct forms of the image ranging from raw pixels, through frequency transformations and the results of principal component and eigenface analyses.

Preliminary work not reported in our experimental results, included approaches to feature selection based on entropy measures. We discovered that subsets of our original $30$ features performed as well using single Gaussian models. An interesting area for future work consists of the integration of a feature selection capability into the full mixture-distance framework.

In this paper we focused on a very restricted setting in which only a single example of each face exists in the database. If instead one assumes the availability of some number of image pairs corresponding to the same person, the task of estimating the parameters of our observation process may be approached more directly. For example, as queries are processed and assuming the machine receives feedback as to whether or not its classification is correct, it might adapt its distance function and one might consider re-formulating the entire framework into a purely on-line setting. A significant message of this paper however is that even in the absence of such feedback, improved distance functions can be found.

Finally we remark that our feature set will most likely limit future gains in accuracy. Variations, however small in 3D pose, camera position and characteristics, and many other sources of error are not explicitly modeled and should be whenever possible. However, forming a conceptual framework towards this end is not nearly as difficult as the associated computational and optimization issues.

Acknowledgments

We thank Johji Tajima and Shizuo Sakamoto of NEC Central Laboratories, Sandy Pentland of the MIT Media Lab, Yael Moses of the Weizmann Institute of Science, and Jonathon Phillips of the Army Research Laboratory for providing the databases we used. The authors acknowledge David W. Jacobs of NEC Research Institute, Sunita L. Hingorani of AT&T Bell Laboratories, and Santhana Krishnamachari of the University of Maryland for their participation in this project's predecessor [3] where portions of Sections 2 and 3 first appeared.

Bibliography

1
R. J. Baron.
Mechanisms of human face recognition.
Int. J. of Man Machine Studies, 15:137-178, 1981.

2
R. Brunelli and T. Poggio.
Face recognition: features versus templates.
IEEE Trans. on PAMI, 15(10):1042-1052, 1993.

3
I. J. Cox, D. W. Jacobs. S. Krishnamachari and P. N. Yianilos.
Experiments on Feature-Based Face Recognition.
NEC Research Institute, 1994.

4
A. J. Goldstein, L. D. Harmon, and A. B. Lesk.
Identification of human faces.
Proc. of the IEEE, 59(5):748-760, 1971.

5
L. D. Harmon, M. K. Khan, R. LAsch, and P. F. Ramig.
Machine identification of human faces.
Pattern Recognition, 13:97-110, 1981.

6
L. D. Harmon, S. C. Kuo, P. F. Ramig, and U. Raudkivi.
Identification of human face profiles by computer.
Pattern Recognition, 10:301-312, 1978.

7
T. Kanade.
Picture processing System By Computer Complex and Recognition of Human Faces.
PhD thesis, Kyoto University, 1973.

8
G. J. Kaufman and K. J. Breeding.
Automatic recognition of human faces from profile silhouettes.
IEEE Trans. on Systems, Man and Cybernetics, SMC-6(2):113-121, 1976.

9
T. Kanade, Computer Recognition of Human Faces.
Birkhäuser Verlag, Stuttgart Germany, 1977.

10
Y. Moses, Y. Adini, and S. Ullman.
Face recognition: the problem of compensating for changes in illumination direction.
In J.-O. Eklundh, editor, Third European Conf. on Computer Vision, pages 286-296, 1994.

11
S. J. Nowlan.
Soft Competitive Adaptation: Neural Network Learning Algorithms based on Fitting Statistical Mixtures.
PhD thesis, School of Computer Science, Carnegie Mellon University, April 1991.

12
A. Pentland, B. Moghaddam, and T. Starner.
View-based and modular eigenspaces for face recognition.
In Computer Vision and Pattern Recognition, pages 84-91, 1994.

13
R. A. Redner and H. F. Walker.
Mixture densities, maximum likelihood and the EM algorithm.
SIAM review, vol. 26, PP. 195-239, 1984.

14
G. D. Riccia and A. Iserles.
Automatic identification of pictures of human faces.
In 1977 Carnahan Conference on Crime Countermeasures, pages 145-148, 1977.

15
S. Sakamoto and J. Tajima.
Face feature analysis for human identification.
Technical report, NEC Central Research Laboratory, 1994.

16
J. Shepherd and H. Ellis.
Face recognition and recall using computer-interactive methods with eye witnesses.
In Processing Images of Faces, editor, V. Bruce and H. Burton, Ablex Publishing, 1992.

17
M. A. Turk and A. P. Pentland.
Face recogntion using eigenfaces.
In IEEE Int. Conf. on Computer Vision and Pattern Recognition, pages 586-591, 1991.

18
S. White.
Features in face recognition algorithms.
Part of the Area Exam for Area II Course VI, MIT, February 1992.

19
J. Tojima and S. Sakamoto.
Private Communication.
1994.

20
L. Xu and M. I. Jordan.
On convergence properties of the em algorithm for gaussian mixtures.
Technical Report A.I. Memo No. 1520, Massachusetts Institute of Technology, Artificial Intelligence Laboratory, January 1995.

21
P. N. Yianilos.
Metric learning via normal mixtures.
Technical report, The NEC Research Institute, Princeton, New Jersey, 1995.

About this document ...

Feature-Based Face Recognition Using Mixture-Distance

This document was generated using the LaTeX2HTML translator Version 2K.1beta (1.47)

Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.

The command line arguments were:
latex2html -split 0 -image_type gif -nonext_page_in_navigation -noprevious_page_in_navigation -up_url ../main.html -up_title 'Publication homepage' -numbered_footnotes faces.tex

The translation was initiated by Peter N. Yianilos on 2002-06-30


Footnotes

... Yianilos1
The first and third authors are with NEC Research Institute, 4 Independence Way, Princeton, NJ 08540. The second is with the University of Montreal, Department of Computer Science. Direct Email to the third author at pny@research.nj.nec.com. This manuscript was completed during October 1995.
... below2
Selection was necessary only because many of the available images were not frontal views.
... problem.3
Formally, the configuration of a parameterized model is just an extension of its parameters.
....4
We use $f=0.99$.


Up: Publication homepage
Peter N. Yianilos
2002-06-30