Showing papers on "Eigenface published in 1997"

Eigenfaces vs. Fisherfaces: recognition using class specific linear projection

Abstract: We develop a face recognition algorithm which is insensitive to large variation in lighting direction and facial expression. Taking a pattern classification approach, we consider each pixel in an image as a coordinate in a high-dimensional space. We take advantage of the observation that the images of a particular face, under varying illumination but fixed pose, lie in a 3D linear subspace of the high dimensional image space-if the face is a Lambertian surface without shadowing. However, since faces are not truly Lambertian surfaces and do indeed produce self-shadowing, images will deviate from this linear subspace. Rather than explicitly modeling this deviation, we linearly project the image into a subspace in a manner which discounts those regions of the face with large deviation. Our projection method is based on Fisher's linear discriminant and produces well separated classes in a low-dimensional subspace, even under severe variation in lighting and facial expressions. The eigenface technique, another method based on linearly projecting the image space to a low dimensional subspace, has similar computational requirements. Yet, extensive experimental results demonstrate that the proposed "Fisherface" method has error rates that are lower than those of the eigenface technique for tests on the Harvard and Yale face databases.

Face recognition : Eigenface, elastic matching, and neural nets : Automated biometrics

Abstract: This paper is a comparative study of three recently proposed algorithms for face recognition: eigenface, autoassociation and classification neural nets, and elastic matching. After these algorithms were analyzed under a common statistical decision framework, they were evaluated experimentally on four individual data bases, each with a moderate subject size, and a combined data base with more than a hundred different subjects. Analysis and experimental results indicate that the eigenface algorithm, which is essentially a minimum distance classifier, works well when lighting variation is small. Its performance deteriorates significantly as lighting variation increases. The elastic matching algorithm, on the other hand, is insensitive to lighting, face position, and expression variations and therefore is more versatile. The performance of the autoassociation and classification nets is upper bounded by that of the eigenface but is more difficult to implement in practice.

Face recognition: eigenface, elastic matching, and neural nets

Abstract: This paper is a comparative study of three recently proposed algorithms for face recognition: eigenface, autoassociation and classification neural nets, and elastic matching. After these algorithms were analyzed under a common statistical decision framework, they were evaluated experimentally on four individual data bases, each with a moderate subject size, and a combined data base with more than a hundred different subjects. Analysis and experimental results indicate that the eigenface algorithm, which is essentially a minimum distance classifier, works well when lighting variation is small. Its performance deteriorates significantly as lighting variation increases. The elastic matching algorithm, on the other hand, is insensitive to lighting, face position, and expression variations and therefore is more versatile. The performance of the autoassociation and classification nets is upper bounded by that of the eigenface but is more difficult to implement in practice.

Parametrized structure from motion for 3D adaptive feedback tracking of faces

Abstract: A real-time system is described for automatically detecting, modeling and tracking faces in 3D. A closed loop approach is proposed which utilizes structure from motion to generate a 3D model of a face and then feed back the estimated structure to constrain feature tracking in the next frame. The system initializes by using skin classification, symmetry operations, 3D warping and eigenfaces to find a face. Feature trajectories are then computed by SSD or correlation-based tracking. The trajectories are simultaneously processed by an extended Kalman filter to stably recover 3D structure, camera geometry and facial pose. Adaptively weighted estimation is used in this filter by modeling the noise characteristics of the 2D image patch tracking technique. In addition, the structural estimate is constrained by using parametrized models of facial structure (eigen-heads). The Kalman filter's estimate of the 3D state and motion of the face predicts the trajectory of the features which constrains the search space for the next frame in the video sequence. The feature tracking and Kalman filtering closed loop system operates at 25 Hz.

Face recognition using range images

Abstract: A system for face recognition using range images as input data is described. The range data acquisition procedure is based on the coded light approach, merging range images that are recorded by two separate sensors. Two approaches, which are known from face recognition based on grey level images have been extended to dealing with range images. These approaches are based on eigenfaces and hidden Markov models, respectively. Experimental results on a database with various range images from 24 persons show very promising results for both recognition methods.

Independent components of face images : A representation for face recognition

Abstract: Methods for obtaining representations of face images based on independent component analysis (ICA) are presented. A global ICA representation is compared to a global representation based on principal component analysis (PCA) for recognizing faces m o s s changes in lighting and changes in pose. For each set of face images, a set of statistically independent source images was found through an unsupervised learning algorithm that maximized the mutual information between the input and the output of a nonlinear transformation (Bell & Sejnowski, 1995). These source images comprised the kernels for the representation. The independent component, kernels gave superior class discriminabiity to the principal component kernels. Recognition across changes in pose with the ICA representation was 93%, compared to 87% with a PCA representation, and across changes in lighting ICA gave 100% correct recognition, compared to 90% with PCA.

Model based faced coding and decoding using feature detection and eigenface coding

Abstract: The method uses a three-dimensional face model and a technique called eigenface decomposition to analyze the video at one end. The facial feature locations and eigenface coding of the face image are sent to a decoder. The decoder synthesizes the face image at the receiving end. Eigenface decoding is used to texture map a three-dimensional model warped by detected feature locations.

A survey of face recognition algorithms and testing results

Abstract: Automated face recognition (AFR) has received increased attention We describe two general approaches to the problem and discuss their effectiveness and robustness with respect to several possible applications We also discuss some issues of run-time performance The AFR technology falls into three main subgroups, which represent more-or-less independent approaches to the problem: neural network solutions, eigenface solutions, and wavelet/elastic matching solutions Each of these first requires that a facial image be identified in a scene, a process called segmentation The image should be normalized to some extent Normalization is usually a combination of linear translation, rotation and scaling, although the elastic matching method includes spatial transformations

Face recognition-matching system effective to images obtained in different imaging conditions

Abstract: A system capable of pattern matching between a less noisy face image and a noisier face image obtained in different imaging conditions. A weighted sum of a covariance matrix calculated from less noisy sample images obtained for a plurality of sample faces and a covariance matrix calculated from differences between the less noisy sample images and corresponding noisier sample images obtained for the sample faces is first obtained. A principal component analysis is applied to the weighted sum to find and store in memory a result with which principal components are calculated from a given face image. In actual pattern matching, a first face image and a second face image are first obtained. The degree of similarity between the first and second face images is calculated based on a conventional formula by using the result of the principal component analysis. The matching between the first and second face images is achieved based on a comparison of the degree of similarity with a predetermined threshold value.

Real-Time Detection of Human Faces in Uncontrolled Environments

Abstract: Mehdi KhosraviNCR Human Interface Technology CenterAtlanta, Georgia, 30309Monson H. HayesGeorgia Institute of Technology, Department of Electrical EngineeringAtlanta, Georgia, 30332ABSTRACTThis paper presents an approach for the detection of human face and eyes in real time and in uncontrolled environments.The system has been implemented on a PC platform with the aid of simple commercial devices such as an NTSC videocamera and a monochrome frame grabber. The approach is based on a probabilistic framework that uses a deformabletemplate model to describe the human face. The system has been tested on both head-and-shoulder sequences as well ascomplex scenes with multiple people and random motion. The system is able to locate the eyes from different head poses(rotations in image plane as well as in depth). The information provided by the location of the eyes is used to extract faceswith frontal pose from a video sequence. The extracted frontal frames can be passed to recognition and classificationsystems for further processing.Keywords : Face Detection, Eye Detection, Face Segmentation, Ellipse Fitting1. INTRODUCTIONIn recent years, face detection from video data has become a popular research area. There are numerous commercialapplications of face detection in face recognition, verification, classification, identification as well as security access andmultimedia. To extract the human faces in an uncontrolled environment most of these applications must deal with thedifficult problems of variations in lighting, variations in pose, occlusion of people by other people, and cluttered or non-uniform backgrounds.A review of the approaches to face detection that have been proposed are described in[1]. In [2], Sung and Poggio presentedan example-based learning approach for locating unoccluded human frontal faces. The approach measures a distancebetween the local image and a few view-based "face" and "non face" pattern prototypes at each image location to locate theface. In [3], Turk and Pentland used the distance to a "face space", defined by "eigenfaces", to locate and track frontalhuman faces. In [4], human faces were detected by searching for significant facial features at each location in the image. In[5]

The Mug-Shot Search Problem

Abstract: Mug-shot search is the classic example of the general problem of searching a large facial image database when starting out with only a mental image of the sought-after face. We have implemented a prototype content-based image retrieval system that integrates composite face creation methods with a face-recognition technique (Eigenfaces) so that a user can both create faces and search for them automatically in a database. Although the Eigenface method has been studied extensively for its ability to perform face identification tasks (in which the input to the system is an on-line facial image to identify), little research has been done to determine how effective it is when applied to the mug shot search problem (in which there is no on-line input image at the outset, and in which the task is similarity retrieval rather than face-recognition). With our prototype system, we have conducted a pilot user study that examines the usefulness of Eigenfaces applied to this problem. The study shows that the Eigenface method, though helpful, is an imperfect model of human perception of similarity between faces. Using a novel evaluation methodology, we have made progress at identifying specific search strategies that, given an imperfect correlation between the system and human similarity metrics, use whatever correlation does exist to the best advantage. The study also indicates that the use of facial composites as query images is advantageous compared to restricting users to database images for their queries.

An eigenfaces-based automatic face recognition system

Abstract: The problem of automatic face recognition (AFR) alone is a difficult task that involves detection and location of faces in a cluttered background, facial feature extraction, subject identification and verification. The main challenge lies in facial feature extraction. This should reduce the intra-person variability (due to changes in geometry, illumination, gesture, and biological changes) and increase the inter-person variability. Various approaches have previously been proposed, including the eigenfaces for which satisfactory experimental results have been reported. The eigenfaces approach assumes that the data is intrinsically low-dimensional. This contribution presents an eigenfaces-based AFR, that guarantees the low-dimensionality assumption by preprocessing steps and multiple eigenspaces. The necessity for pre-processing steps has already been recognized by other groups. In this paper, the need for multiple eigenspaces and the corresponding operative criterion is established.

Object recognition with luminance, rotation and location invariance

Abstract: We propose a neural network based on image synthesis, histogram adaptive quantization and the discrete cosine transformation (DCT) for object recognition with luminance, rotation and location invariance. An efficient representation of the invariant features is constructed using a three-dimensional memory structure. The performance of luminance and rotation invariance is illustrated by reduced error rates in face recognition. The error rate of using a two-dimensional DCT is improved from 13.6% to 2.4% with the aid of the proposed image synthesis procedure. The 2.4% error rate is better than all previously reported results using Karhunen-Loeve (1990) transform convolution networks and eigenface models. In using the DCT, our approach also enjoys the additional advantage of greatly reduced computational complexity.

Face recognition using mixture-distance and raw images

Abstract: Earlier work suggests that mixture-distance can improve the performance of feature-based face recognition systems in which only a single training example is available for each individual. In this work we investigate the non-feature-based eigenfaces technique of Turk and Pentland (1991), replacing Euclidean distance with mixture-distance. In mixture-distance, a novel distance function is constructed based on local second-order statistics as estimated by modeling the training data with a mixture of normal densities. The approach is described and experimental results on a database of 600 people are presented, showing that mixture-distance can reduce the error rate by up to 73.9%. In the experimental setting considered, the results indicate that the simplest form of mixture distance yields considerable improvement. Additional, but less dramatic, improvement was possible with more complex forms. The results show that even in the absence of multiple training examples for each class, it is sometimes possible to infer an improved distance function from a statistical model of the training data. Therefore, researchers using eigenfaces or similar pattern recognition techniques may find significant advantages by considering alternative distance metrics such as mixture-distance.

Lip-shape Dependent Face Verification

Abstract: We present a multimodal approach to face verification which draws on two distinct knowledge sources of biometric information concerning the subject. A B-spline lip tracker provides the control information regarding the state of the lip shape which is used by a conventional eigenface based face verification system to confirm or reject a claimed personal identity. The performance of the system tested on the M2VTS database shows a promising improvement over the unimodal approach. This improvement derives from the achieved reduction in the population entropy of the models, thus minimising the probability of imposter acceptance.

Multiresolution eigenface-components

Abstract: This paper presents a face recognition system that imitates the multiresolution processing technique employed by the human visual system. In the proposed system, a different degree of importance is assigned to each part of a face image, and each region of the face image is processed with a different resolution. This proposed system reduces the computational complexity of the eigenface method, and achieves higher compression ratios and higher recognition rates in comparison with the eigenface method. Experimental results are presented and discussed.

Face recognition using statistical models

Abstract: We describe the use of flexible models for the representation of shape and grey-level appearance of human faces. The models are controlled by a small number of parameters, which can be used to code the overall appearance of a face for image compression and classification. Shape and grey-level appearance are included in a single model. Discriminant analysis allows the isolation of variation important for classification of identity. We have performed both face recognition and face synthesis experiments and present the results in this paper.

Dual eigenspace method for human face recognition

Abstract: The authors present an effective scheme called the dual eigenspace method (DEM) for automated face recognition. Based on the K-L transform, the dual eigenspaces are constructed by extracting algebraic features of training samples and applied to face identification with a two-layer minimium distance classifier. Experimental results show that DEM is significantly better than the traditional eigenface method (TEM).

A DCT-based adaptive metric learning model using asymptotic local information measure

Abstract: We present an adaptive metric learning vector quantization procedure based on the discrete-cosine transform (DCT) for accurate face recognition used in multimedia applications. Since the set of learning samples may be small, we employ a mixture model of prior distributions. The model selection method, which minimizes the cross entropy between the real distribution and the modeled one, is presented to optimize the mixture number and local metric parameters. The structural risk minimization is used to facilitate an asymptotic approximation of the cross entropy for models of fixed complexity. We also provide a formula to estimate the model complexity derived from the minimum description length criterion. The structural risk minimization method proposed achieves an recognition error rate of 2.29% using the ORL database, which is better than previously reported numbers using the Karhunen-Loeve transform convolution network, the hidden Markov model and the eigenface model.

Shape normalisation for face recognition

Abstract: This paper presents methods for shape normalisation of face images. Localisation and shape normalisation are prerequisites for face recognition algorithms like the eigenface approach. Other established face recognition methods like labeled graph matching can also be accelerated by providing normalised face image databases.

A principal component based BDNN for face recognition

Abstract: We propose a high performance two-stage hybrid structure for face recognition. The first stage is an eigenface based recognizer, which serves as a candidate faces selector. From our experience, the Top 1 recognition rate is only 65%, however the Top 10 hit rate can be up to 98.15%. The Top 10 candidate faces are similar to each other, thus these faces are called simial faces. Since the projections of the similar faces are too close in the eigenspace, it's very hard to distinguish a target face from similar face set. Thus, we propose the "horizontal average gray scale (HAGS)" as a new type of feature for the second stage recognizer. A paired-Bayesian-decision neural network (pBDNN) is used for the second stage recognizer, which identifies the target from the similar faces. Supported by the proposed feature, a pDBNN could make an accurate classification between any two similar faces. In order to demonstrate the proposed hybrid system, we conducted some experiments on an in house database, which contains 675 images taken from 135 people. The training data for the pBDNN were small orientation (-22.5/spl deg/, 0/spl deg/, 22.5/spl deg/), and the large orientation (-45/spl deg/ and 45/spl deg/) images were for testing. Our experimental results show that the hybrid recognition structure improves the recognition rate for 17% more than the eigenface system alone (65%) without any rejection, and 26% more with 31% of rejection.

On Relation of Image Compression and Image Association

Abstract: Based on two image compression schemes (MIT and RNS), it is shown that it is possible to associate similar object images using their intermediate representation. Thus both methods can be applied to large image database for both goals: high quality image compression and reliable search for queries by image content. MIT scheme of Moghaddam and Pentland is specialized to face images. It moves image comparison task from high dimensional image space to low dimensional principal subspace spanned on eigenfaces. The closest point in the subspace is used for image association. RNS scheme of the author represents images (not limited to a certain scene type) by recurrent neural subnetworks which together with a competition layer create an associative memory. The single recurrent subnetwork N i is designed for the i-th image and it implements a stochastic nonlinear operator F i. It can be shown that under realistic assumptions F i has a unique attractor which is located in the vicinity of the original image. When at the input a noisy, incomplete or distorted image is presented, the associative recall is implemented in two stages. Firstly, a competition layer finds the most invariant subnetwork. Next, the selected recurrent subnetwork reconstructs in few iterations the original image.