Tangent Distance (TD) is one classical method for invariant pattern classification. However, conventional TD need pre-obtain tangent vectors, which is difficult except for image objects. This paper extends TD to more general pattern classification tasks. The basic assumption is that tangent vectors can be approximately represented by the pattern variations. We propose three probabilistic subspace models to encode the variations: the linear subspace, nonlinear subspace, and manifold subspace models. These three models are addressed in a unified view, namely Probabilistic Tangent Subspace (PTS). Experiments show that PTS can achieve promising classification performance in non-image data sets.

Kernel PCA is an efficient method for nonlinear feature extraction. We address two issues in kernel PCA based feature extraction and classification. First, it extracts features without utilizing sample label information. Second, it does not provide a practical means to choose the dimensionality for principal subspace. In this paper, one kind of side-information is incorporated into kernel PCA to solve the first problem. And a complete probabilistic density function is estimated in kernel space so that the choice of dimensionality for principal subspace becomes less important. The proposed model is named probabilistic kernel feature subspace (PKFS). Experiments show that it achieves promising performance and outperforms many other algorithms in classification.

We study a pattern in the fingerprint called a crease, a kind of stripe which irregularly crosses the normal fingerprint patterns (ridges and valleys). Creases will cause spurious minutiae when using a conventional feature detection algorithm, and therefore decreases the recognition rate of fingerprint identification. By representing the crease using a parameterized rectangle, we design an optimal filter as a detector. We employ a multi-channel filtering framework to detect creases in different orientations. In each channel, PCA is used to extract a rectangle's parameters from the raw detected results. Our algorithm is demonstrated by experiments.

True minutiae extraction in fingerprint image is critical to the performance of an automated identification system. Generally, a set of endings and bifurcations (both called feature points) can be obtained by the thinning image from which the true minutiae of the fingerprint are extracted by using the rules based on the structure of ridges. However, considering some false and true minutiae have similar ridge structures in the thinning image, in a lot of cases, we have to explore their difference in the binary image or the original gray image. In this paper, we first define the different types of feature points and analyze the properties of their ridge structures in both thinning and binary images for the purpose of distinguishing the true and false minutiae. Based on the knowledge of these properties, a fingerprint post-processing approach is developed to eliminate the false minutiae and at the same time improve the thinning image for further application. Many experiments are performed and the results have shown the great effectiveness of the approach.

We present a fully automated face recognition system (FAFRS) which contains two components: eyes detection and face recognition. Firstly, a hybrid neural method is proposed to localize human eyes. Then, a dual eigenspaces method (DEM) is developed to extract algebraic features from normalized facial images and perform the recognition task. Experimental results show that the proposed system is effective and robust under different conditions.

This paper presents a novel stroke tracing approach to recovering the strokes from the handwriting images. It is carried out not only based on pixel by pixel, but considering the local relation of handwriting strokes. Firstly the primary skeletons of handwriting are produced through the binarization and thinning. Secondly, after the reliable skeletons and unreliable skeletons are classified based on the width information of handwriting, the strokes segments are obtained by the connecting and tracing on the reliable skeletons. Finally, for making the recovered strokes more natural, the post-processing of tracing is discussed. Experimental results show the effectiveness of the approach.

A new vector quantization image coding scheme called orientation normalized vector quantizer (ONVQ) is presented. During the encoding process, the orientation field of an image is first calculated and recorded using a quad-tree data structure; then, the image is partitioned into digitalized round blocks. These round blocks are normalized to the horizontal direction according to their original orientation and the vector which is quantized. The decoder obtains reproduction vectors from a codebook, and restores them to their original orientations. Since blocks of different orientations can use the same codeword, ONVQ needs a much smaller codebook; therefore, low bit rates are achieved (0.6-0.8 bpp for 256/spl times/256 images). The subjective quality of ONVQ-coded images is more satisfying than that obtained by a simple vector quantization scheme.>

A method is described for the color image description by using stable view points of an image obtained after Gaussian smoothing. The basic idea is to determine the view points and corresponding viewing fields at various levels from the color information and define the local image by using these points and fields so that the image matching can be realized. This approach may be developed into a generalized method for image recognition.>