Songcan Chen

Robust image segmentation using FCM with spatial constraints based on new kernel-induced distance measure

Fuzzy c-means clustering (FCM) with spatial constraints (FCM/spl I.bar/S) is an effective algorithm suitable for image segmentation. Its effectiveness contributes not only to the introduction of fuzziness for belongingness of each pixel but also to exploitation of spatial contextual information. Although the contextual information can raise its insensitivity to noise to some extent, FCM/spl I.bar/S still lacks enough robustness to noise and outliers and is not suitable for revealing non-Euclidean structure of the input data due to the use of Euclidean distance (L/sub 2/ norm). In this paper, to overcome the above problems, we first propose two variants, FCM/spl I.bar/S/sub 1/ and FCM/spl I.bar/S/sub 2/, of FCM/spl I.bar/S to aim at simplifying its computation and then extend them, including FCM/spl I.bar/S, to corresponding robust kernelized versions KFCM/spl I.bar/S, KFCM/spl I.bar/S/sub 1/ and KFCM/spl I.bar/S/sub 2/ by the kernel methods. Our main motives of using the kernel methods consist in: inducing a class of robust non-Euclidean distance measures for the original data space to derive new objective functions and thus clustering the non-Euclidean structures in data; enhancing robustness of the original clustering algorithms to noise and outliers, and still retaining computational simplicity. The experiments on the artificial and real-world datasets show that our proposed algorithms, especially with spatial constraints, are more effective.

Sparsity preserving projections with applications to face recognition

Dimensionality reduction methods (DRs) have commonly been used as a principled way to understand the high-dimensional data such as face images. In this paper, we propose a new unsupervised DR method called sparsity preserving projections (SPP). Unlike many existing techniques such as local preserving projection (LPP) and neighborhood preserving embedding (NPE), where local neighborhood information is preserved during the DR procedure, SPP aims to preserve the sparse reconstructive relationship of the data, which is achieved by minimizing a L1 regularization-related objective function. The obtained projections are invariant to rotations, rescalings and translations of the data, and more importantly, they contain natural discriminating information even if no class labels are provided. Moreover, SPP chooses its neighborhood automatically and hence can be more conveniently used in practice compared to LPP and NPE. The feasibility and effectiveness of the proposed method is verified on three popular face databases (Yale, AR and Extended Yale B) with promising results.

A novel kernelized fuzzy C-means algorithm with application in medical image segmentation

Image segmentation plays a crucial role in many medical imaging applications. In this paper, we present a novel algorithm for fuzzy segmentation of magnetic resonance imaging (MRI) data. The algorithm is realized by modifying the objective function in the conventional fuzzy C-means (FCM) algorithm using a kernel-induced distance metric and a spatial penalty on the membership functions. Firstly, the original Euclidean distance in the FCM is replaced by a kernel-induced distance, and thus the corresponding algorithm is derived and called as the kernelized fuzzy C-means (KFCM) algorithm, which is shown to be more robust than FCM. Then a spatial penalty is added to the objective function in KFCM to compensate for the intensity inhomogeneities of MR image and to allow the labeling of a pixel to be influenced by its neighbors in the image. The penalty term acts as a regularizer and has a coefficient ranging from zero to one. Experimental results on both synthetic and real MR images show that the proposed algorithms have better performance when noise and other artifacts are present than the standard algorithms.

Recognizing partially occluded, expression variant faces from single training image per person with SOM and soft k-NN ensemble

Most classical template-based frontal face recognition techniques assume that multiple images per person are available for training, while in many real-world applications only one training image per person is available and the test images may be partially occluded or may vary in expressions. This paper addresses those problems by extending a previous local probabilistic approach presented by Martinez, using the self-organizing map (SOM) instead of a mixture of Gaussians to learn the subspace that represented each individual. Based on the localization of the training images, two strategies of learning the SOM topological space are proposed, namely to train a single SOM map for all the samples and to train a separate SOM map for each class, respectively. A soft k nearest neighbor (soft k-NN) ensemble method, which can effectively exploit the outputs of the SOM topological space, is also proposed to identify the unlabeled subjects. Experiments show that the proposed method exhibits high robust performance against the partial occlusions and variant expressions.

Locality preserving CCA with applications to data visualization and pose estimation

Canonical correlation analysis (CCA) is a major linear subspace approach to dimensionality reduction and has been applied to image processing, pose estimation and other fields. However, it fails to discover or reveal the nonlinear correlation relationship between two sets of features. In contrast, its kernelized nonlinear version, KCCA, can overcome such a shortcoming, but the global kernelization of CCA restrains KCCA itself from effectively discovering the local structure of the data with complex and nonlinear characteristics. Recently, the locality methods, such as locally linear embedding (LLE) and locality preserving projections (LPP), are proposed to discover the low dimensional manifold embedded into the original high dimensional space. Compared to the subspace based methods, these locality methods take into account the local neighborhood structure of the data, and can discover the intrinsic structure of data to a better degree, which benefits subsequent computation. Inspired by the locality based methods, in this paper, we incorporate such an idea into CCA and propose locality preserving CCA (LPCCA) to discover the local manifold structure of the data and further apply it to data visualization and pose estimation. In addition, a fast algorithm of LPCCA is proposed for some special cases. The experiments show that LPCCA can both capture the intrinsic structure characteristic of the given data and achieve higher pose estimation accuracy than both CCA and KCCA.

Subpattern-based principle component analysis

Abstract We propose a subpattern-based principle component analysis (SpPCA). The traditional PCA operates directly on a whole pattern represented as a vector and acquires a set of projection vectors to extract global features from given training patterns. SpPCA operates instead directly on a set of partitioned subpatterns of the original pattern and acquires a set of projection sub-vectors for each partition to extract corresponding local sub-features and then synthesizes them into global features for subsequent classification. The experimental results show that the proposed SpPCA has (much) better classification performances on all the real-life benchmark datasets than PCA.

Constraint Score: A new filter method for feature selection with pairwise constraints

Feature selection is an important preprocessing step in mining high-dimensional data. Generally, supervised feature selection methods with supervision information are superior to unsupervised ones without supervision information. In the literature, nearly all existing supervised feature selection methods use class labels as supervision information. In this paper, we propose to use another form of supervision information for feature selection, i.e. pairwise constraints, which specifies whether a pair of data samples belong to the same class (must-link constraints) or different classes (cannot-link constraints). Pairwise constraints arise naturally in many tasks and are more practical and inexpensive than class labels. This topic has not yet been addressed in feature selection research. We call our pairwise constraints guided feature selection algorithm as Constraint Score and compare it with the well-known Fisher Score and Laplacian Score algorithms. Experiments are carried out on several high-dimensional UCI and face data sets. Experimental results show that, with very few pairwise constraints, Constraint Score achieves similar or even higher performance than Fisher Score with full class labels on the whole training data, and significantly outperforms Laplacian Score.

Letters: Adaptively weighted sub-pattern PCA for face recognition

Adaptively weighted Sub-pattern PCA (Aw-SpPCA) for face recognition is presented in this paper. Unlike PCA based on a whole image pattern, Aw-SpPCA operates directly on its sub-patterns partitioned from an original whole pattern and separately extracts features from them. Moreover, unlike both SpPCA and mPCA that neglect different contributions made by different parts of the human face in face recognition, Aw-SpPCA can adaptively compute the contributions of each part and then endows them to a classification task in order to enhance the robustness to both expression and illumination variations. Experiments on three standard face databases show that the proposed method is competitive.

Enhanced (PC) 2 A for face recognition with one training image per person

Abstract Recently, a method called (PC) 2 A was proposed to deal with face recognition with one training image per person. As an extension of the standard eigenface technique, (PC) 2 A combines linearly each original face image with its corresponding first-order projection into a new face and then performs principal component analysis (PCA) on a set of the newly combined (training) images. It was reported that (PC) 2 A could achieve higher accuracy than the eigenface technique through using 10–15% fewer eigenfaces. In this paper, we generalize and further enhance (PC) 2 A along two directions. In the first direction, we combine the original image with its second-order projections as well as its first-order projection in order to acquire more information from the original face, and then similarly apply PCA to such a set of the combined images. In the second direction, instead of combining them, we still regard the projections of each original image as single derived images to augment training image set, and then perform PCA on all the training images available, including the original ones and the derived ones. Experiments on the well-known FERET database show that the enhanced versions of (PC) 2 A are about 1.6–3.5% more accurate and use about 47.5–64.8% fewer eigenfaces than (PC) 2 A.

Making FLDA applicable to face recognition with one sample per person

In face recognition, the Fisherface approach based on Fisher linear discriminant analysis (FLDA) has obtained some success. However, FLDA fails when each person just has one training face sample available because of nonexistence of the intra-class scatter. In this paper, we propose to partition each face image into a set of sub-images with the same dimensionality, therefore obtaining multiple training samples for each class, and then apply FLDA to the set of newly produced samples. Experimental results on the FERET face database show that the proposed approach is feasible and better in recognition performance than E(PC) 2 A.