Xiangyan Zeng

Multi-class feature selection for texture classification

In this paper, a multi-class feature selection scheme based on recursive feature elimination (RFE) is proposed for texture classifications. The feature selection scheme is performed in the context of one-against-all least squares support vector machine classifiers (LS-SVM). The margin difference between binary classifiers with and without an associated feature is used to characterize the discriminating power of features for the binary classification. A new criterion of min-max is used to mix the ranked lists of binary classifiers for multi-class feature selection. When compared to the traditional multi-class feature selection methods, the proposed method produces better classification accuracy with fewer features, especially in the case of small training sets.

Milestones in electron crystallography

Electron crystallography determines the structure of membrane embedded proteins in the two-dimensionally crystallized state by cryo-transmission electron microscopy imaging and computer structure reconstruction. Milestones on the path to the structure are high-level expression, purification of functional protein, reconstitution into two-dimensional lipid membrane crystals, high-resolution imaging, and structure determination by computer image processing. Here we review the current state of these methods. We also created an Internet information exchange platform for electron crystallography, where guidelines for imaging and data processing method are maintained. The server (http://2dx.org) provides the electron crystallography community with a central information exchange platform, which is structured in blog and Wiki form, allowing visitors to add comments or discussions. It currently offers a detailed step-by-step introduction to image processing with the MRC software program. The server is also a repository for the 2dx software package, a user-friendly image processing system for 2D membrane protein crystals.

Texture representation based on pattern map

We propose a pixel-pattern-based texture feature (PPBTF) which is insensitive to variance of illumination. A gray scale image is transformed into a pattern map where edges and lines (bars) to be used for characterizing the texture information are classified by pattern matching. Based on the pattern map, a feature vector is comprised of the numbers of the pixels belonging to each pattern. We use the image basis functions obtained by principal component analysis as the templates for pattern matching. The characteristics of the feature is comprehensively analyzed through an application to texture image segmentation. Comparisons with multichannel filtering methods show that PPBTF feature is quite time saving and free of the influence of illumination.

3D reconstruction from 2D crystal image and diffraction data.

Electron crystallography of 2D protein crystals can determine the structure of membrane embedded proteins at high resolution. Images or electron diffraction patterns are recorded with the electron microscope of the frozen hydrated samples, and the 3D structure of the proteins is then determined by computer data processing. Here we introduce the image-processing algorithms for crystallographic Fourier space based methods using the Medical Research Council (MRC) programs, and illustrate the usage of the software packages 2dx, XDP, and IPLT.

Signal separation by independent component analysis based on a genetic algorithm

We propose a genetic algorithm for blind source separation (BSS). The BSS problem is to obtain the independent components of original source signals from mixed signals. The original sources that are mutually independent and are mixed linearly by an unknown matrix are retrieved by a separating procedure using independent component analysis (ICA). The goal of ICA is to find a separating matrix so that the separated signals are as independent as possible. Many neural learning algorithms for minimizing the dependency among signals have been proposed for obtaining the separating matrix. The effectiveness of these algorithms, however, is affected by the neuron activation functions that depend on the probability distribution of the signals. In our method, the separating matrix is evolved by a genetic algorithm (GA) that does not need activation functions and works on an evolutionary mechanism. The kurtosis that is a simple and original criterion for independence is used in the fitness function of GA. The applicability of the proposed method for blind source separation is demonstrated by the simulation results.

Single Particle 3D Reconstruction for 2D Crystal Images of Membrane Proteins

In cases where ultra-flat cryo-preparations of well-ordered two-dimensional (2D) crystals are available, electron crystallography is a powerful method for the determination of the high-resolution structures of membrane and soluble proteins. However, crystal unbending and Fourier-filtering methods in electron crystallography three-dimensional (3D) image processing are generally limited in their performance for 2D crystals that are badly ordered or non-flat. Here we present a single particle image processing approach, which is implemented as an extension of the 2D crystallographic pipeline realized in the 2dx software package, for the determination of high-resolution 3D structures of membrane proteins. The algorithm presented, addresses the low single-to-noise ratio (SNR) of 2D crystal images by exploiting neighborhood correlation between adjacent proteins in the 2D crystal. Compared with conventional single particle processing for randomly oriented particles, the computational costs are greatly reduced due to the crystal-induced limited search space, which allows a much finer search space compared to classical single particle processing. To reduce the considerable computational costs, our software features a hybrid parallelization scheme for multi-CPU clusters and computer with high-end graphic processing units (GPUs). We successfully apply the new refinement method to the structure of the potassium channel MloK1. The calculated 3D reconstruction shows more structural details and contains less noise than the map obtained by conventional Fourier-filtering based processing of the same 2D crystal images.

Edge detection and texture segmentation based on independent component analysis

We present a new feature extraction technique based on independent component analysis (ICA). We use ICA to learn the basis junctions of natural images and then the basis functions are used as pattern templates for feature detections. The successful applications of the proposed method to edge detection and texture segmentation are demonstrated.

Image processing of 2D crystal images.

Electron crystallography of membrane proteins uses cryo-transmission electron microscopy to image frozen-hydrated 2D crystals. The processing of recorded images exploits the periodic arrangement of the structures in the images to extract the amplitudes and phases of diffraction spots in Fourier space. However, image imperfections require a crystal unbending procedure to be applied to the image before evaluation in Fourier space. We here describe the process of 2D crystal image unbending, using the 2dx software system.

Image feature representation by the subspace of nonlinear PCA

In subspace pattern recognition, the basis vectors represent the features of the data and define the class. In the previous works, the standard principal component analysis is used to derive the basis vectors. Compared with the standard PCA, a nonlinear PCA can provide the high-order statistics and result in non-orthogonal basis vectors. We combine a nonlinear PCA and a subspace classifier to extract the edge and line features in an image. The simulation results indicate that the basis vectors from the nonlinear PCA can classify the edge patterns better than those from a linear PCA.

Feature Selection Using Recursive Feature Elimination for Handwritten Digit Recognition

In this paper, a new feature selection method with applications to handwritten digit recognition is proposed. This method is based on recursive feature elimination (RFE) in least squares support vector machines (LS-SVM). Digit recognition is achieved by one-against-all LS-SVMs. The RFE method is adapted to multi-class classification in two ways. One is to prune features for each binary LS-SVM classifier independently, and the other is to prune features for all the binary classifiers jointly. The multi-class RFE is also compared with the wrapper feature selection method which uses genetic algorithms. The experimental results indicate that the joint pruning algorithm yields the best performance and selects more features relevant to intrinsic characteristics of digits.