Search

While face detection seems a solved problem under general conditions, most state-of-the-art systems degrade rapidly when faces are partially occluded by other objects. This paper presents a solution to detect partially occluded faces by reasonably modifying the AdaBoost-based face detector. Our basic idea is that the weak classifiers in the AdaBoost-based face detector, each corresponding to a Haar-like feature, are inherently a patch-based model. Therefore, one can divide the whole face region into multiple patches, and map those weak classifiers to the patches. The weak classifiers belonging to each patch are re-formed to be a new classifier to determine if it is a valid face patch - without occlusion. Finally, we combine all of the valid face patches by assigning the patches with different weights to make the final decision whether the input subwindow is a face. The experimental results show that the proposed method is promising for the detection of occluded faces

Scientists find that the human perception is based on the similarity on the manifold of data set. Isometric feature mapping (isomap) is one of the representative techniques of manifold. It is intuitive, well understood and produces reasonable mapping results. However, if the input data for manifold learning are corrupted with noises, the isomap algorithm is topologically unstable. In this paper, we present an improved manifold learning method when the input data are images - the image Euclidean distance based isomap (imisomap), in which we use a new distance for images called image Euclidean distance (IMED). Experimental results demonstrate a consistent performance improvement of the algorithm imisomap over the traditional isomap based on Euclidean distance

Kernel Fisher discriminant analysis (KFDA) has achieved great success in pattern recognition recently. However, the training process of KFDA is too time consuming (even intractable) for a large training set, because, for a training set with n examples, both its between-class and within-class scatter matrices are of n times n and the time complexity of the KFDA training process is of O(n3). Aiming at this problem, this paper employs bagging technique to decrease the time-space cost of KFDA training process. In addition, this paper is more than just a simple application of bagging. We have made an important adaptation which can further guarantee the performance of KFDA. Our experimental results demonstrate that the proposed method can not only greatly reduce the cost of time of the training process, but also achieve higher recognition accuracy than traditional KFDA and the simple application of bagging

Viewpoint variance is one of the inevitable problems in vision based sign language recognition. However, most researchers avoid this problem by assuming a special view, especially the front view. In the paper, we propose a verification method for viewpoint invariant sign language recognition. In general, there are two major variances between two video sequences of the same sign: performance variance and viewpoint variance. For small performance variance, DTW can help us eliminate it. When there is only viewpoint variance between two sequences, we can consider the two sequences as obtained synchronously by a stereo vision system. Thus, for the current input, we can judge whether the known template is the matched one by verifying whether the two sequences can be considered as obtained by a stereo vision system. Our experiments demonstrate the efficiency of the proposed method. Furthermore, such verification method can be easily extended to other recognition tasks.

This paper proposes a novel simple yet effective generative model based on Local Visual Primitives (LVP) for face modeling and classification. The LVPs, as the pattern of local face region, are learnt by clustering a great number of local patches. Visually, these LVPs correspond to intuitive low-level micro visual structures very well, and they are expected to constitute those high-level semantic features, such as eyes, nose and mouth. We show that, though face appearances vary dramatically, these LVPs are very effective for face image reconstruction. For face recognition, block-based histograms of the LVPs indexes are extracted as the face representation to compare for classification. Primary experiments on FERET face database have shown that the LVP method can achieve encouraging recognition rate.

This paper present a novel method for robust illumination-tolerant face recognition based on the Gabor phase and a probabilistic similarity measure. Invited by the work in Eigenphases of J. Lou, et al. (2002) by using the phase spectrum of face images, we use the phase information of the multi-resolution and multi-orientation Gabor filters. We show that the Gabor phase has more discriminative information and it is tolerate to illumination variations. Then we use a probabilistic similarity measure based on a Bayesian (MAP) analysis of the difference between the Gabor phases of two face images. We train the model using some images in the illumination subset of CMU-PIE database and test on the other images of CMU-PIE database and the Yale B database and get comparative results

In this paper, we introduce a novel discriminative feature which is efficient for pose estimation. The multi-view face representation is based on Local Gabor Binary Patterns( LGBP) and encodes the local facial characteristics in to a compact feature histogram. In LGBP, Gabor filters can extract the feature of the orientation of head and Local Binary Pattern(LBP) can extract the features of facial local orientation. To keep the spatial information of the multi-view face images, LGBP is operated on many subregions of the images. The combination of them can represent well and truly the multi-view face images. Considering the derived feature space, a radial basis function(RBF) kernel SVM classifier is trained to estimate pose. Extensive experiments demonstrate that the facial representation can be effective for pose estimation. is a face. The experimental results show that the proposed method is promising for the detection of occluded faces.

Face representations based on Gabor features have achieved great success in face recognition, such as Elastic Graph Matching, Gabor Fisher Classifier (GFC), and AdaBoosted Gabor Fisher Classifier (AGFC). In GFC and AGFC, either down-sampled or selected Gabor features are analyzed in holistic mode by a single classifier. In this paper, we propose a novel patch-based GFC (PGFC) method, in which Gabor features are spatially partitioned into a number of patches, and on each patch one GFC is constructed as component classifier to form the final ensemble classifier using sum rule. The positions and sizes of the patches are learned from a training data using AdaBoost. Experiments on two large-scale face databases (FERET and CAS-PEAL-R1) show that the proposed PGFC with only tens of patches outperforms the GFC and AGFC impressively.

A novel background generation method based on non-parametric background model is presented for background subtraction. We introduce a new model, named as effect components description (ECD), to model the variation of the background, by which we can relate the best estimate of the background to the modes (local maxima) of the underlying distribution. Based on ECD, an effective background generation method, most reliable background mode (MRBM), is developed. The basic computational module of the method is an old pattern recognition procedure, the mean shift, which can be used recursively to find the nearest stationary point of the underlying density function. The advantages of this method are three-fold: first, backgrounds can be generated from image sequence with cluttered moving objects; second, backgrounds are very clear without blur effect; third, it is robust to noise and small vibration. Extensive experimental results illustrate its good performance

We present a novel method to enhance training set for face detection with nonlinearly generated examples from the original data. The motivation is from Support Vector Machines (SVM) that, for classification problems, examples lying close to class boundary usually have more influence and thus are more informative than those far from the boundary. We utilize a nonlinear technique - reduced set (RS) method and a new image distance metric to generate new examples, and then add them to the original collected database to enhance it. Extensive experiments show that the proposed approach has an encouraging performance.

Spatial histogram* of Local Binary Pattern (LBP) and Local Gabor Binary Pattern (LGBP) has been successfully applied to face recognition and achieved state-of-the-art performance. Both LBP and LGBP utilize traditional histogram matching method such as histogram intersection for face classification. In this paper, we propose a statistical extension for L(G)BP similarity computation by introducing Fisher Discriminant Analysis (FDA) of the L(G)BP spatial histogram ?features?. More than a simple application of FDA, we have constructed Ensemble of Piecewise FDA (EPFDA) classifiers, each of which is designed using one segment of the entire spatial histogram features. We show that this extension not only greatly reduces the feature dimension but also brings very impressive performance improvement. Especially, we have made a large step to recognizing all the faces in the standard FERET face database.