Search

Hitherto, one major challenge to sign language recognition is how to develop approaches that scale well with increasing vocabulary size. In large vocabulary speech recognition realm, it is effective to use phonemes instead of words as the basic units. This idea can be used in large vocabulary Sign Language recognition, too. In this paper, Etyma are defined to be the smallest unit in a sign language, that is, a unit that has some meaning and distinguishes one sign from the others. They can be seen as phonemes in Sign Language. Two approaches to large vocabulary Chinese Sign Languagerecognition are discussed in this paper. One uses etyma and the other uses whole signs as the basic units. Two CyberGloves and a Pohelmus 3-D tracker with three receivers positioned on the wrist of CyberGlove and the back are used as input device. Etymon- and word- based recognition systems are introduced, which are designed to recognize 2439 etyma and 5100 signs. And then the experimental results of these two systems are given and analyzed. [ABSTRACT FROM AUTHOR]

In Sign Language recognition, one of the problems is to collect enough data. Data collection for both training and testing is a laborious but necessary step. Almost all of the statistical methods used in Sign Language Recognition suffer from this problem. Inspired by the crossover and mutation of genetic algorithms, this paper presents a method to enlarge Chinese Sign language database through re-sampling from existing sign samples. Two initial samples of the same sign are regarded as parents. They can reproduce their children by crossover. To verify the effectiveness of the proposed method, some experiments are carried out on a vocabulary with 350 static signs. Each sign has 4 samples. Three samples are used to be the original generation. These three original samples and their offspring are used to construct the training set, and the remaining sample is used for testing. The experimental results show that the data generated by the proposed method are effective. [ABSTRACT FROM AUTHOR]

This paper proposes the AdaBoost Gabor Fisher Classifier (AGFC) for robust face recognition, in which a chain AdaBoost learning method based on Bootstrap re-sampling is proposed and applied to face recognition with impressive recognition performance. Gabor features have been recognized as one of the most successful face representations, but it is too high dimensional for fast extraction and accurate classification. In AGFC, AdaBoost is exploited to select optimally the most informative Gabor features (hereinafter as AdaGabor features). The selected low-dimensional AdaGabor features are then classified by Fisher discriminant analysis for final face identification. Our experiments on two large-scale face databases, FERET and CAS-PEAL (with 5789 images of 1040 subjects), have shown that the proposed method can effectively reduce the dimensionality of Gabor features and greatly increase the recognition accuracy. In addition, our experimental results show its robustness to variations in facial expression and accessories. [ABSTRACT FROM AUTHOR]

The development of web and digital camera nowadays has made it easier to collect more than hundreds of thousands of examples. How to train a face detector based on the collected enormous face database? This paper presents a manifold-based method to subsample. That is, we learn the manifold from the collected face database and then subsample training set by the estimated geodesic distance which is calculated during the manifold learning. Using the subsampled training set based on the manifold, we train an AdaBoost-based face detector. The trained detector is tested on the MIT+CMU frontal face test set. The experimental results show that the proposed method is effective and efficient to train a classifier confronted with the huge database. [ABSTRACT FROM AUTHOR]

A novel nonlinear discriminant analysis method, Kernelized Decision Boundary Analysis (KDBA), is proposed in our paper, whose Decision Boundary feature vectors are the normal vector of the optimal Decision Boundary in terms of the Structure Risk Minimization principle. We also use a simple method to prove a property of Support Vector Machine (SVM) algorithm, which is combined with the optimal Decision Boundary Feature matrix to make our method consistent with the Kernel Fisher method(KFD). Moreover, KDBA is easily used in its applications, and the traditional Decision Boundary Analysis implementations are computationally expensive and sensitive to the size of the problem. Text classification problem is first used to testify the effectiveness of KDBA. Then experiments on the large-scale face database, the CAS-PEAL database, have illustrated its excellent performance compared with some popular face recognition methods such as Eigenface, Fisherface, and KFD. Keywords: Face Recognition, Kernel Fisher, Support Vector Machine [ABSTRACT FROM AUTHOR]

Pose and illumination changes from picture to picture are two main barriers toward full automatic face recognition. In this paper, a novel method to handle both pose and lighting condition simultaneously is proposed, which calibrates the pose and lighting condition to a pre-set reference condition through an illumination invariant 3D face reconstruction. First, some located facial landmarks and a priori statistical deformable 3D model are used to recover an elaborate 3D shape. Based on the recovered 3D shape, the "texture image" calibrated to a standard illumination is generated by spherical harmonics ratio image and finally the illumination independent 3D face is reconstructed completely. The proposed method combines the strength of statistical deformable model to describe the shape information and the compact representations of the illumination in spherical frequency space, and handle both the pose and illumination variation simultaneously. This algorithm can be used to synthesize virtual views of a given face image and enhance the performance of face recognition. The experimental results on CMU PIE database show that this method can significantly improve the accuracy of the existed face recognition method when pose and illumination are inconsistent between gallery and probe sets. [ABSTRACT FROM AUTHOR]

This paper presents a novel approach to face recognition, named Multi-resolution Histograms of Local Variation Patterns (MHLVP), in which face images are represented as the concatenation of the local spatial histogram of local variation patterns computed from the multi-resolution Gabor features. For a face image with abundant texture and shape information, aGabor feature map(GFM) is computed by convolving the image with each of the forty multi-scale and multi-orientation Gabor filters. Each GFM is then divided into small non-overlapping regions to enhance its shape information, and then Local Binary Pattern (LBP) histograms are extracted for each region and concatenated into a feature histogram to enhance the texture information in the specific GFM. Further more, all of the feature histograms extracted from the forty GFMs are further concatenated into a single feature histogram as the final representation of the given face image. Eventually, the identification is achieved by histogram intersection operation. Our experimental results on FERET face databases show that the proposed method performs terrifically better than the performance of some classical results including the best results in FERET'97. [ABSTRACT FROM AUTHOR]