Yangsheng Wang

Generalized quotient image

We present a unified framework for modeling intrinsic properties of face images for recognition. It is based on the quotient image (QI) concept, in particular on the existing works of QI, spherical harmonic, image ratio and retinex. Under this framework, we generalize these previous works into two new algorithms: (1) non-point light quotient image (NPL-QI) extends QI to deal with non-point light sources by modeling non-point light directions using spherical harmonic bases; (2) self-quotient image (S-QI) extends QI to perform illumination subtraction without the need for alignment and no shadow assumption. Experimental results show that our algorithms can significantly improve the performance of face recognition under varying illumination conditions.

Technical Section: Fitting 3D garment models onto individual human models

Designing an elegant 3D virtual garment model for a 3D virtual human model is labor-intensive, because most existing garment models are custom-made for a specific human model and cannot be easily reused for other individuals. In this paper, we propose a novel method for fitting a given 3D garment model onto human models of various body shapes and poses. The fitting is accomplished by deforming the garment mesh to match the shapes of the human models by using a combination of the following: skeleton-driven volumetric deformation, garment-human shape similarity matching and evaluation, the constraints of garment-human contact, and garment-human ease allowance. Experiments show that our approach performs very well and has the potential to be used in the garment design industry.

Gabor feature selection for face recognition using improved adaboost learning

Though AdaBoost has been widely used for feature selection and classifier learning, many of the selected features, or weak classifiers, are redundant. By incorporating mutual information into AdaBoost, we propose an improved boosting algorithm in this paper. The proposed method fully examines the redundancy between candidate classifiers and selected classifiers. The classifiers thus selected are both accurate and non-redundant. Experimental results show that the strong classifier learned using the proposed algorithm achieves a lower training error rate than AdaBoost. The proposed algorithm has also been applied to select discriminative Gabor features for face recognition. Even with the simple correlation distance measure and 1-NN classifier, the selected Gabor features achieve quite high recognition accuracy on the FERET database, where both expression and illumination variance exists. When only 140 features are used, the selected features achieve as high as 95.5% accuracy, which is about 2.5% higher than that of features selected by AdaBoost.

Learning with Cascade for Classification of Non-Convex Manifolds

Images of a visual object, such as human face, reside in a complicated manifold in the high dimensional image space, when the object is subject to variations in pose, illumination, and other factors. Viola and Jones have successfully tackled difficult nonlinear classification problem for face detection using AdaBoost learning. Moreover, their simple-to-complex cascade of classifiers structure makes the learning and classification even more effective. While training with cascade has been used effectively in many works [4, 5, 6, 7, 2, 3, 8, 9, 10], an understanding of the role of the cascade strategy is still lacking. In this paper, we analyze the problem of classifying non-convex manifolds using AdaBoost learning with and without using cascade. We explain that the divide-and-conquer strategy in cascade learning has a great contribution on learning a complex classifier for non-convex manifolds. We prove that AdaBoost learning with cascade is effective when a complete or over-complete set of features (or weak classifiers) is available. Experiments with both synthesized and real data demonstrate that AdaBoost learning with cascade leads to improved convergence and accuracy.

Iris image capture system design for personal identification

Iris image acquisition is a key issue in iris recognition, as the quality of the captured image greatly affects the performance of the overall system This paper first discusses the current status of iris capture devices and then describes the design of a new iris sensor Experimental results with the iris images captured using the new iris image acquisition device are also presented in this paper.

Statistical Learning of Evaluation Function for ASM/AAM Image Alignment

Alignment between the input and target objects has great impact on the performance of image analysis and recognition system, such as those for medical image and face recognition. Active Shape Models (ASM) [1] and Active Appearance Models (AAM) [2, 3] provide an important framework for this task. However, an effective method for the evaluation of ASM/AAM alignment results has been lacking. Without an alignment quality evaluation mechanism, a bad alignment cannot be identified and this can drop system performance.

Face decorating system based on improved active shape models

This paper presents a face decorating system, which can do makeup on a face image, such as wearing glass, beard or lipstick. In the framework, an improved face alignment method is proposed to localize the key landmarks of a face, which would be used to locate decorations. Active Shape Models(ASMs), as a robust image alignment method is employed to localize such landmarks. In this work, the authors review the conventional ASMs algorithm for face alignment, and present several improvements on it. Its believed that traditional ASMs is heavily dependent on initial states and prone to local minima. To improve the stability as well as its efficiency, much work is done. First, the eyes are roughly localized in the face area, which are used to initialize the shape model and evaluate the result. Then conventional point distribution model(PDM) is replaced by a newly proposed combined PDM. Experiments on a database containing 200 labelled face images show that the proposed method performs significantly better than traditional ASMs. Finally, the improved method was used to implement a face decorating system.

3D object modelling for entertainment applications

Recent advances in three-dimensional (3D) data acquisition techniques have offered an alternative to the traditional 2D metamorphosis (or morphing) approaches, which gradually change a source object through intermediate objects into a target object. In this paper, we approach 3D metamorphosis via a novel 3D modelling technique, which reconstructs a fairly complex object with a single B-Spline patch. Our object representation is compact - over 90% compression rate can be achieved. Despite such huge amount of data reduction, our method achieves similar rendering result to that using polygonal representation. Our approach also allows a one-to-one mapping from the object space to a common parameter space to be established, to allow automatic correspondence between a pair of objects. This way to establishing object correspondence is advantageous over the common connectivity generation process, with which, if either the source or target object is changed, the whole process of establishing correspondences must be repeated. Several aesthetically pleasing examples of 3D morphing are demonstrated using the proposed method.

Mesh smoothing via adaptive bilateral filtering

In this paper, we present an adaptive bilateral filtering algorithm that can be used to remove unavoidable noise from 3D mesh data generated by initial stages. Selecting the parameters for bilateral filters automatically, this algorithm smoothes meshes in the normal field using anisotropic character of local neighborhood triangles. Experimental results demonstrate that the proposed method remove light noise from meshes and reserve fine features of meshes as good as best results of other methods, with the advantage of none user-assisted parameters setting. Visual comparisons display that the method proposed in this paper performs better than other smoothing method for heavy noisy mesh.

Evaluation of face alignment solutions using statistical learning

We propose a statistical learning approach for constructing an evaluation function for face alignment. A nonlinear classification function is learned from a set of positive (good alignment) and negative (bad alignment) training examples to effectively distinguish between qualified and un-qualified alignment results. The AdaBoost learning algorithm is used, where weak classifiers are constructed based on edge features and combined into a strong classifier. Several strong classifiers are learned in stages using bootstrap samples during the training. The evaluation function thus learned gives a quantitative confidence and the good-bad classification is achieved by comparing the confidence with a learned optimal threshold. We point out the importance of using cascade strategy in the stagewise learning of strong classifiers. The divide-and-conquer strategy not only dramatically increases the speed of classification, but also makes the training easier and the good-bad classification more effective. Experimental results demonstrate that the classification function learned using the proposed approach provides semantically more meaningful scoring than the reconstruction error used in AAM for classification between qualified and un-qualified face alignment.