Skip to main content
main-content

Über dieses Buch

The two-volume set LNCS 7951 and 7952 constitutes the refereed proceedings of the 10th International Symposium on Neural Networks, ISNN 2013, held in Dalian, China, in July 2013. The 157 revised full papers presented were carefully reviewed and selected from numerous submissions. The papers are organized in following topics: computational neuroscience, cognitive science, neural network models, learning algorithms, stability and convergence analysis, kernel methods, large margin methods and SVM, optimization algorithms, varational methods, control, robotics, bioinformatics and biomedical engineering, brain-like systems and brain-computer interfaces, data mining and knowledge discovery and other applications of neural networks.

Inhaltsverzeichnis

Frontmatter

Computational Neuroscience and Cognitive Science

Information Transfer Characteristic in Memristic Neuromorphic Network

Memristive nanodevices can support exactly the same learning function as spike-timing-dependent plasticity in neuroscience, and thus the exploration for the evolution and self-organized computing of memristor-based neuromorphic networks becomes reality. We mainly study the STDP-driven refinement effect on memristor-based crossbar structure and its information transfer characteristic. The results show that self-organized refinement could enhance the information transfer of memristor crossbar, and the dependence of memristive device on current direction and the balance between potentiation and depression are of crucial importance. This gives an inspiration for resolving the power consumption issue and the so called sneak path problem.

Quansheng Ren, Qiufeng Long, Zhiqiang Zhang, Jianye Zhao

Generation and Analysis of 3D Virtual Neurons Using Genetic Regulatory Network Model

Neuronal morphology is significant for understanding structure-function relationships and brain information processing in computational neuroscience. So it is very important to simulate neuronal morphology completely and accurately. In this paper, we present a novel approach for efficient generation of 3D virtual neurons using genetic regulatory network model. This approach describes dendritic geometry and topology by locally inter-correlating morphological variables which can be represented by the dynamics of gene expression. The experimental results show that the generating virtual neurons that are anatomically indistinguishable and accurate from experimentally traced real neurons.

Xianghong Lin, Zhiqiang Li

A Finite-Time Convergent Recurrent Neural Network Based Algorithm for the L Smallest k-Subsets Sum Problem

For a given set

S

of

n

real numbers, a

k

-subset means a subset of

k

distinct elements of

S

. It is obvious that there are totally

$C_{n}^{k}$

different combinations. The

L

smallest

k

-subsets sum problem is defined as finding

L

k

-subsets whose summation of subset elements are the

L

smallest among all possible combinations. This problem has many applications in research and the real world. However the problem is very computationally challenging. In this paper, a novel algorithm is proposed to solve this problem. By expressing all the

$C_{n}^{k}$

k

-subsets with a network, the problem is converted to finding the

L

shortest loopless paths in this network. By combining the

L

shortest paths algorithm and the finite-time convergent recurrent neural network, a new algorithm for the

L

smallest

k

-subsets problem is developed. And experimental results show that the proposed algorithm is very effective and efficient.

Shenshen Gu

Spike Train Pattern and Firing Synchronization in a Model of the Olfactory Mitral Cell

In the olfactory system, both the temporal spike structure and spatial distribution of neuronal activity are important for processing odor information. In this paper, a biophysically-detailed, spiking neuronal model is used to simulate the activity of olfactory bulb. It is shown that by varying some key parameters such as maximal conductances of

Ks

,

Nap

and

Na

, the spike train of single neuron can exhibit various firing patterns. In the olfactory bulb, synchronization in coupled neurons is also investigated as the coupling strength gets increased. Synchronization process can be identified by correlation coefficient and phase plot. It is illustrated that the coupled neurons can exhibit different types of synchronization when the coupling strength increases. These results may be instructive to understand information transmission in olfactory system.

Ying Du, Rubin Wang, Jingyi Qu

Efficiency Improvements for Fuzzy Associative Memory

FAM is an Associative Memory that uses operators of Fuzzy Logic and Mathematical Morphology (MM). FAMs possess important advantages including noise tolerance, unlimited storage, and one pass convergence. An important property, deciding FAM performance, is the ability to capture contents of each pattern, and associations of patterns. Standard FAMs capture either contents or associations of patterns well, but not both of them. In this paper, we propose a novel FAM that effectively stores both contents and associations of patterns. We improve both learning and recalling processes of FAM. In learning process, the associations and contents are stored by mean of input and output patterns and they are generalised by erosion operator. In recalling process, a new threshold is added to output function to improve outputs. Experiments show that noise tolerance of the proposed FAM is better than standard FAMs with different types of noise.

Nong Thi Hoa, The Duy Bui, Trung Kien Dang

A Study of Neural Mechanism in Emotion Regulation by Simultaneous Recording of EEG and fMRI Based on ICA

The combination of electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) is a very attractive aim in neuroscience, in order to achieve both high temporal and spatial resolution for the non-invasive study of cognitive brain function. In this paper, we record simultaneous EEG–fMRI of the same subject in emotional processing experiment in order to explore the characteristics of different emotional picture processing, and try to find the difference of the subject’ brain hemisphere when viewing different valence emotional pictures. For fMRI data, we study the participant’s brain active region, and examine related blood oxygen level—dependent(BOLD) response. For EEG data, we focus on the amplitude of the late positive potential (LPP). We find that the amplitude of the LPP correlated significantly with BOLD intensity in visual cortex and amygdala, prefrontal is also modulated by different picture categories.

Tiantong Zhou, Hailing Wang, Ling Zou, Renlai Zhou, Nong Qian

Emotion Cognitive Reappraisal Research Based on Simultaneous Recording of EEG and BOLD Responses

The combination of electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) is a very attractive aim in neuroscience, in order to achieve both high temporal and spatial resolution for the non-invasive study of cognitive brain function. In this paper, we record simultaneous EEG–fMRI of three subjects, study the participants’ brain active regions, examine related blood oxygen level—dependent(BOLD) response for fMRI data, and focus on the effects of reappraisal instructions on the amplitude of the late positive potential(LPP) for EEG data. We find that emotion cognitive reappraisal result in early prefrontal cortex responses, decrease negative emotion experience and amygdala response. Besides, the study indicates that reappraisal decrease the magnitude of the LPP.

Ling Zou, Yi Zhang, Lin Yuan, Nong Qian, Renlai Zhou

Convergence of Chaos Injection-Based Batch Backpropagation Algorithm For Feedforward Neural Networks

This paper considers the convergence of chaos injection-based backpropagation algorithm. Both the weak convergence and strong convergence results are theoretically established.

Huisheng Zhang, Xiaodong Liu, Dongpo Xu

Discovering the Multi-neuronal Firing Patterns Based on a New Binless Spike Trains Measure

In this paper, we proposed a method which presented a new definition of different multi-step interval ISI-distance distribution of single neuronal spike trains and formed a new feature vector to represent the original spike trains. It is a binless spike train’s measure method. We used spectral clustering algorithm on new multi-dimensional feature vectors to detect the multiple neuronal firing patterns. We tested this method on standard data set in machine learning, neuronal surrogate data set and in vivo multi-electrode recordings respectively. Results shown that the method proposed in this paper can effectively improve the clustering accuracy in standard data set and detect the firing patterns in neuronal spike trains.

Hu Lu, Hui Wei

A Study on Dynamic Characteristics of the Hippocampal Two-Dimension Reduced Neuron Model under Current Conductance Changes

In the paper, based on the computer simulation, the hippocampal two-dimension reduced neuron model is taken as the object, and its dynamic bifurcation characteristics are analyzed and discussed in detail by the neurodynamic analysis methods. When the maximum conductance of the instantaneous sodium channel and the maximum conductance of the delay-rectified potassium channel are changed, the neuron model undergoes the supercritical Andronov-Hopf bifurcation from the rest state to the continuous discharge state. The neuron model is a resonator with the monostable state and has the common dynamics of the resonator. This investigation is helpful to know and investigate deeply the dynamic characteristics and the bifurcation mechanism of the hippocampal neuron by the computer simulation.

Yueping Peng, Xinxu Wang, Xiaohua Qiu

Neural Network Models, Learning Algorithms, Stability and Convergence Analysis

Overcoming the Local-Minimum Problem in Training Multilayer Perceptrons with the NRAE-MSE Training Method

The normalized risk-averting error (NRAE) training method presented in ISNN 2012 is capable of overcoming the local-minimum problem in training neural networks. However, the overall success rate is unsatisfactory. Motivated by this problem, a modification, called the NRAE-MSE training method is herein proposed. The new method trains neural networks with respect to NRAE with a fixed

λ

in the range of 10

6

-10

11

, and takes excursions to train with the standard mean squared error (MSE) from time to time. Once an excursion produces a satisfactory MSE with cross-validation, the entire NRAE-MSE training stops. Numerical experiments show that the NRAE-MSE training method has a success rate of 100% in all the testing examples each starting with a large number of randomly selected initial weights.

James Ting-Ho Lo, Yichuan Gui, Yun Peng

Generalized Single-Hidden Layer Feedforward Networks

In this paper, we propose a novel generalized single-hidden layer feedforward network (GSLFN) by employing polynomial functions of inputs as output weights connecting randomly generated hidden units with corresponding output nodes. The main contributions are as follows. For arbitrary

N

distinct observations with

n

-dimensional inputs, the augmented hidden node output matrix of the GSLFN with

L

hidden nodes using any infinitely differentiable activation functions consists of

L

sub-matrix blocks where each includes

n

 + 1 column vectors. The rank of the augmented hidden output matrix is proved to be no less than that of the SLFN, and thereby contributing to higher approximation performance. Furthermore, under minor constraints on input observations, we rigorously prove that the GLSFN with

L

hidden nodes can exactly learn

L

(

n

 + 1) arbitrary distinct observations which is

n

 + 1 times what the SLFN can learn. If the approximation error is allowed, by means of the optimization of output weight coefficients, the GSLFN may require less than

N

/(

n

 + 1) random hidden nodes to estimate targets with high accuracy. Theoretical results of the GSLFN evidently perform significant superiority to that of SLFNs.

Ning Wang, Min Han, Guifeng Yu, Meng Joo Er, Fanchao Meng, Shulei Sun

An Approach for Designing Neural Cryptography

Neural cryptography is widely considered as a novel method of exchanging secret key between two neural networks through mutual learning. This paper puts forward a generalized architecture to provide an approach to designing novel neural cryptography. Meanwhile, by taking an in-depth investigation on the security of neural cryptography, a heuristic rule is proposed. These results can effectively guide us to designing secure neural cryptography. Finally, an example is given to demonstrate the effectiveness of the proposed structure and the heuristic rule.

Nankun Mu, Xiaofeng Liao

Bifurcation of a Discrete-Time Cohen-Grossberg-Type BAM Neural Network with Delays

A tri-neuron discrete-time Cohen-Grossberg BAM neural network with delays is investigated in this paper. By analyzing the corresponding characteristic equations, the asymptotical stability of the null solution and the existence of Neimark-Sacker bifurcations are discussed. By applying the normal form theory and the center manifold theorem, the direction of the Neimark-Sacker bifurcation and the stability of bifurcating periodic solutions are obtained. Numerical simulations are given to illustrate the obtained results.

Qiming Liu

Stability Criteria for Uncertain Linear Systems with Time-Varying Delay

This paper is concerned with the stability criteria for uncertain systems with time-varying delay. The parameter uncertainties are supposed to be norm-bounded. By using Lyapunov functional and integral inequality, some delay-dependent stability criteria are obtained. Numerical examples are given to demonstrate the effectiveness of proposed method.

Huimin Liao, Manchun Tan, Shuping Xu

Generalized Function Projective Lag Synchronization between Two Different Neural Networks

The generalized function projective lag synchronization (GFPLS) is proposed in this paper. The scaling functions which we have investigated are not only depending on time, but also depending on the networks. Based on Lyapunov stability theory, a feedback controller and several sufficient conditions are designed such that the response networks can realize lag-synchronize with the drive networks. Finally, the corresponding numerical simulations are performed to demonstrate the validity of the presented synchronization method.

Guoliang Cai, Hao Ma, Xiangqian Gao, Xianbin Wu

Application of Local Activity Theory of CNN to the Coupled Autocatalator Model

The study of chemical reactions with oscillating kinetics has drawn increasing interest over the last few decades. However the dynamical properties of the coupled nonlinear dynamic system are difficult to deal with. The local activity principle of the Cellular Nonlinear Network (CNN) introduced by Chua has provided a powerful tool for studying the emergence of complex behaviors in a homogeneous lattice formed by coupled cells. Based on the Autocatalator Model introduced by Peng.B, this paper establishes a two dimensional coupled Autocatalator CNN system. Using the analytical criteria for the local activity calculates the chaos edge of the Autocatalator CNN system. The numerical simulations show that the emergence may exist if the selected cell parameters are nearby the edge of chaos domain. The Autocatalator CNN can exhibit periodicity and chaos.

Guangwu Wen, Yan Meng, Lequan Min, Jing Zhang

Passivity Criterion of Stochastic T-S Fuzzy Systems with Time-Varying Delays

In this paper, the passivity for stochastic Takagi-Sugeno (T-S) fuzzy systems with time-varying delays is investigated without assuming the differentiability of the time-varying delays. By utilizing the Lyapunov functional method, the It

$\hat{o}$

differential rule and the matrix inequality techniques, a delay-dependent criterion to ensure the passivity for T-S fuzzy systems with time-varying delays is established in terms of linear matrix inequalities (LMIs) that can be easily checked by using the standard numerical software.

Zhenjiang Zhao, Qiankun Song

Parallel Computation of a New Data Driven Algorithm for Training Neural Networks

Different from some early learning algorithms such as backpropagation (BP) or radial basis function (RBF) algorithms, a new data driven algorithm for training neural networks is proposed. The new data driven methodology for training feedforward neural networks means that the system modeling are performed directly using the input-output data collected from real processes, To improve the efficiency, the parallel computation method is introduced and the performance of parallel computing for the new data driven algorithm is analyzed. The results show that, by using the parallel computing mechanisms, the training speed can be much higher.

Daiyuan Zhang

Stability Analysis of a Class of High Order Fuzzy Cohen-Grossberg Neural Networks with Mixed Delays and Reaction-Diffusion Terms

In this paper, we investigate a class of high order fuzzy Cohen-Grossberg neural networks (HOFCGNN) with mixed delays which include time variable dalay and unbounded delays. Based on the properties of M-matrix, by constructing vector Lyapunov functions and applying differential inequalities, the sufficient conditions ensuring existence, uniqueness, and global exponential stability of the equilibrium point of HOFCGNN with mixed delays and reaction-diffusion terms are obtained.

Weifan Zheng, Jiye Zhang, Mingwen Wang

A Study on the Randomness Reduction Effect of Extreme Learning Machine with Ridge Regression

In recent years, Extreme Learning Machine (ELM) has attracted comprehensive attentions as a universal function approximator. Comparing to other single layer feedforward neural networks, its input parameters of hidden neurons can be randomly generated rather than tuned, and thereby saving a huge amount of computational power. However, it has been pointed out that the randomness of ELM parameters would result in fluctuating performances. In this paper, we intensively investigate the randomness reduction effect by using a regularized version of ELM, named Ridge ELM (RELM). Previously, RELM has been shown to achieve generally better generalization than the original ELM. Furthermore, we try to demonstrate that RELM can also greatly reduce the fluctuating performance with 12 real world regression tasks. An insight into this randomness reduction effect is also given.

Meng Joo Er, Zhifei Shao, Ning Wang

Stability of Nonnegative Periodic Solutions of High-Ordered Neural Networks

In this paper, a class of high-ordered neural networks are investigated. By rigorous analysis, a set of sufficient conditions ensuring the existence of a nonnegative periodic solution and its

$R^n_+$

-asymptotical stability are established. The results obtained can also be applied to the first-ordered neural networks.

Lili Wang, Tianping Chen

Existence of Periodic Solution for Competitive Neural Networks with Time-Varying and Distributed Delays on Time Scales

In this paper, under the condition without assuming the boundedness of the activation functions, the competitive neural networks with time-varying and distributed delays are studied. By means of contraction mapping principle, the existence and uniqueness of periodic solution are investigated on time scales.

Yang Liu, Yongqing Yang, Tian Liang, Xianyun Xu

Global Exponential Stability in the Mean Square of Stochastic Cohen-Grossberg Neural Networks with Time-Varying and Continuous Distributed Delays

In this paper, the global exponential stability in the mean square of stochastic Cohen-Grossberg neural networks (SCGNNS) with mixed delays is studied. By applying the Lyapunov function, stochastic analysis technique and inequality techniques, some sufficient conditions are obtained to ensure the exponential stability in the mean square of the SCGNNS. An example is given to illustrate the theoretical results.

Tian Liang, Yongqing Yang, Manfeng Hu, Yang Liu, Li Li

A Delay-Partitioning Approach to Stability Analysis of Discrete-Time Recurrent Neural Networks with Randomly Occurred Nonlinearities

This paper considers the problem of stability analysis for discrete-time recurrent neural networks with randomly occurred nonlinearities (RONs) and time-varying delay. By utilizing new Lyapunov-Krasovskii functions and delay-partitioning technique, the stability criteria are proposed in terms of linear matrix inequality (LMI). We have also shown that the conservatism of the conditions is a non-increasing function of the number of delay partitions. A numerical example is provided to demonstrate the effectiveness of the proposed approach.

Jianmin Duan, Manfeng Hu, Yongqing Yang

The Universal Approximation Capabilities of Mellin Approximate Identity Neural Networks

Universal approximation capability of feedforward neural networks with one hidden layer states that these networks are dense in the space of functions. In this paper, the concept of the Mellin approximate identity functions is proposed. By using this concept, It is shown that feedforward Mellin approximate identity neural networks with one hidden layer can approximate any positive real continuous function to any degree of accuracy. Moreover, universal approximation capability of these networks is extended to positive real Lebesgue spaces.

Saeed Panahian Fard, Zarita Zainuddin

H  ∞  Filtering of Markovian Jumping Neural Networks with Time Delays

This paper focuses on studying the filtering problem of Markovian jumping neural networks with time delays. Based on a stochastic Lyapunov functional, a delay-dependent design criterion is presented under which the resulting filtering error system is stochastically stable and a prescribed

H

 ∞ 

performance is guaranteed. It is shown that the gain matrices of the desired filter and the optimal performance index are simultaneously obtained by handing a convex optimization problem subject to some coupled linear matrix inequalities, which can be efficiently solved by some standard algorithms.

He Huang, Xiaoping Chen, Qiang Hua

Convergence Analysis for Feng’s MCA Neural Network Learning Algorithm

The minor component analysis is widely used in many fields, such as signal processing and data analysis, so it has very important theoretical significance and practical values for the convergence analysis of these algorithms. In this paper we seek the convergence condition for Feng’s MCA learning algorithm in deterministic discrete time system. Finally numerical experiments show the correctness of our theory.

Zhengxue Li, Lijia You, Mingsong Cheng

Anti-periodic Solutions for Cohen-Grossberg Neural Networks with Varying-Time Delays and Impulses

In this paper, we discuss the existence and exponential stability of the anti-periodic solution for delayed Cohen-Grossberg neural networks with impulsive effects. First we give some sufficient conditions to ensure existence and stability of the anti-periodic solutions. Then we present an example with numerical simulations to illustrate our results.

Abdujelil Abdurahman, Haijun Jiang

Global Robust Exponential Stability in Lagrange Sense for Interval Delayed Neural Networks

The problem of global robust exponential stability in Lagrange sense for the interval delayed neural networks (IDNNs) with general activation functions is investigated. Based on the Lyapunov stability, a differential inequality and linear matrix inequalities (LMIs) technique, some conditions to guarantee the IDNNs global exponential stability in Lagrange sense are provided. Meanwhile, the specific estimation of globally exponentially attractive sets of the addressed system are also derived. Finally, a numerical example is provided to illustrate the effectiveness of the method proposed.

Xiaohong Wang, Xingjun Chen, Huan Qi

The Binary Output Units of Neural Network

When solving a multi-classification problem with k kinds of samples, if we use a multiple linear perceptron, k output nodes will be widely-used. In this paper, we introduce binary output units of multiple linear perceptron by analyzing the classification problems of vertices of the regular hexahedron in the Three-dimensional Euclidean Space. And we define Binary Approach and One-for-Each Approach to the problem. Then we obtain a theorem with the help of which we can find a Binary Approach that requires more less classification planes than the One-for-Each Approach when solving a One-for-Each Separable Classification Problem. When we apply the Binary Approach to the design of output units of multiple linear perceptron, the output units required will decrease greatly and more problems could be solved.

Qilin Sun, Yan Liu, Zhengxue Li, Sibo Yang, Wei Wu, Jiuwu Jin

Kernel Methods, Large Margin Methods and SVM

Support Vector Machine with Customized Kernel

In the past two decades, Support Vector Machine (SVM) has become one of the most famous classification techniques. The optimal parameters in an SVM kernel are normally obtained by cross validation, which is a time-consuming process. In this paper, we propose to learn the parameters in an SVM kernel while solving the dual optimization problem. The new optimization problem can be solved iteratively as follows:

(a) Fix the parameters in an SVM kernel; solve the variables

α

i

in the dual optimization problem.

(b) Fix the variables

α

i

; solve the parameters in an SVM kernel by using the Newton-Raphson method.

It can be shown that (a) can be optimized by using standard methods in training the SVM, while (b) can be solved iteratively by using the Newton-Raphson method. Experimental results conducted in this paper show that our proposed technique is feasible in practical pattern recognition applications.

Guangyi Chen, Tien Dai Bui, Adam Krzyzak, Weihua Liu

Semi-supervised Kernel Minimum Squared Error Based on Manifold Structure

Kernel Minimum Squared Error (KMSE) has been receiving much attention in data mining and pattern recognition in recent years. Generally speaking, training a KMSE classifier, which is a kind of supervised learning, needs sufficient labeled examples. However, there are usually a large amount of unlabeled examples and few labeled examples in real world applications. In this paper, we introduce a semi-supervised KMSE algorithm, called

Laplacian regularized KMSE

(LapKMSE), which explicitly exploits the manifold structure. We construct a

p

nearest neighbor graph to model the manifold structure of labeled and unlabeled examples. Then, LapKMSE incorporates the structure information of labeled and unlabeled examples in the objective function of KMSE by adding a Laplacian regularized term. As a result, the labels of labeled and unlabeled examples vary smoothly along the geodesics on the manifold. Experimental results on several synthetic and real-world datasets illustrate the effectiveness of our algorithm.

Haitao Gan, Nong Sang, Xi Chen

Noise Effects on Spatial Pattern Data Classification Using Wavelet Kernel PCA

A Monte Carlo Simulation Study

The kernel-based feature extraction method is of importance in applications of artificial intelligence techniques to real-world problems. It extends the original data space to a higher dimensional feature space and tends to perform better in many non-linear classification problems than a linear approach. This work makes use of our previous research outcomes on the construction of wavelet kernel for kernel principal component analysis (KPCA). Using Monte Carlo simulation approach, we study noise effects of the performance of wavelet kernel PCA in spatial pattern data classification. We investigate how the classification accuracy change when feature dimension is changed. We also compare the classification accuracy obtained from the single-scale and multi-scale wavelet kernels to demonstrate the advantage of using multi-scale wavelet kernel in KPCA. Our study show that multi-scale wavelet kernel performs better than single-scale wavelet kernel in classification of data that we consider. It also demonstrates the usefulness of multi-scale wavelet kernels in application of feature extraction in kernel PCA.

Shengkun Xie, Anna T. Lawniczak, Sridhar Krishnan

SVM-SVDD: A New Method to Solve Data Description Problem with Negative Examples

Support Vector Data Description(SVDD) is an important method to solve data description or one-class classification problem. In original data description problem, only positive examples are provided in training. The performance of SVDD can be improved when a few negative examples are available which is known as SVDD_neg. Intuitively, these negative examples should cause an improvement on performance than SVDD. However, the performance of SVDD may become worse when some negative examples are available. In this paper, we propose a new approach “SVM-SVDD”, in which Support Vector Machine(SVM) helps SVDD to solve data description problem with negative examples efficiently. SVM-SVDD obtains its solution by solving two convex optimization problems in two steps. We show experimentally that our method outperforms SVDD_neg in both training time and accuracy.

Zhigang Wang, Zeng-Shun Zhao, Changshui Zhang

Applying Wavelet Packet Decomposition and One-Class Support Vector Machine on Vehicle Acceleration Traces for Road Anomaly Detection

Road condition monitoring through real-time intelligent systems has become more and more significant due to heavy road transportation. Road conditions can be roughly divided into normal and anomaly segments. The number of former should be much larger than the latter for a useable road. Based on the nature of road condition monitoring, anomaly detection is applied, especially for pothole detection in this study, using accelerometer data of a riding car. Accelerometer data were first labeled and segmented, after which features were extracted by wavelet packet decomposition. A classification model was built using one-class support vector machine. For the classifier, the data of some normal segments were used to train the classifier and the left normal segments and all potholes were for the testing stage. The results demonstrate that all 21 potholes were detected reliably in this study. With low computing cost, the proposed approach is promising for real-time application.

Fengyu Cong, Hannu Hautakangas, Jukka Nieminen, Oleksiy Mazhelis, Mikko Perttunen, Jukka Riekki, Tapani Ristaniemi

Aeroengine Turbine Exhaust Gas Temperature Prediction Using Process Support Vector Machines

The turbine exhaust gas temperature (EGT) is an important parameter of the aeroengine and it represents the thermal health condition of the aeroengine. By predicting the EGT, the performance deterioration of the aeroengine can be deduced in advance and its remaining time-on-wing can be estimated. Thus, the flight safety and the economy of the airlines can be guaranteed. However, the EGT is influenced by many complicated factors during the practical operation of the aeroengine. It is difficult to predict the change tendency of the EGT effectively by the traditional methods. To solve this problem, a novel EGT prediction method named process support vector machine (PSVM) is proposed. The solving process of the PSVM, the kernel functional construction and its parameter optimization are also investigated. Finally, the proposed prediction method is utilized to predict the EGT of some aeroengine, and the results are satisfying.

Xu-yun Fu, Shi-sheng Zhong

The Effect of Lateral Inhibitory Connections in Spatial Architecture Neural Network

Based on the theories of lateral inhibition and artificial neural network (ANN), the different lateral inhibitory connections among the hidden neurons of SANN are studied. With the connect mode of activation-inhibition-activation, the SANN will obtain a higher learning accuracy and generalization ability. Furthermore, this inhibitory connection considers both the activation before and after been inhibited by surrounding neurons. The effectiveness of this inhibitory mode is demonstrated by simulation results.

Gang Yang, Jun-fei Qiao, Wei Li, Wei Chai

Empirical Mode Decomposition Based LSSVM for Ship Motion Prediction

An empirical mode decomposition (EMD) based Lease square support vector machines (LSSVM) is proposed for ship motion prediction. For this purpose, the original ship motion series were first decomposed into several intrinsic mode functions (IMFs), then a LSSVM model was used to model each of the extracted IMFs, so that the tendencies of these IMFs could be accurately predicted. Finally, the prediction results of all IMFs are combined to formulate an output for the original ship motion series. Experiments on chaotic datasets and real ship motion data are used to test the effectiveness of the proposed algorithm.

Zhou Bo, Shi Aiguo

Optimization Algorithms / Variational Methods

Optimal Calculation of Tensor Learning Approaches

Most algorithms have been extended to the tensor space to create algorithm versions with direct tensor inputs. However, very unfortunately basically all objective functions of algorithms in the tensor space are non-convex. However, sub-problems constructed by fixing all the modes but one are often convex and very easy to solve. However, this method may lead to difficulty converging; iterative algorithms sometimes get stuck in a local minimum and have difficulty converging to the global solution. Here, we propose a computational framework for constrained and unconstrained tensor methods. Using our methods, the algorithm convergence situation can be improved to some extent and better solutions obtained. We applied our technique to Uncorrelated Multilinear Principal Component Analysis (UMPCA), Tensor Rank one Discriminant Analysis (TR1DA) and Support Tensor Machines (STM); Experiment results show the effectiveness of our method.

Kai Huang, Liqing Zhang

Repeatable Optimization Algorithm Based Discrete PSO for Virtual Network Embedding

Aiming at reducing the link load and improving substrate network resource utilization ratio, we model the virtual network embedding (VNE) problem as an integer linear programming and present a discrete particle swarm optimization based algorithm to solve the problem. The approach allows multiple virtual nodes of the same VN can be embedded into the same physical node as long as there is enough resource capacity. It not only can cut down embedding processes of virtual link and reduce the embedding time, but also can save the physical link cost and make more virtual networks to be embedded at the same time. Simulation results demonstrate that comparing with the existing VNE algorithm, the proposed algorithm performs better for accessing more virtual networks and reducing embedding cost.

Ying Yuan, Cui-Rong Wang, Cong Wan, Cong Wang, Xin Song

An Energy-Efficient Coverage Optimization Method for the Wireless Sensor Networks Based on Multi-objective Quantum-Inspired Cultural Algorithm

The energy-efficiency coverage of wireless sensor network is measure by the network cover rate and the node redundancy rate. To solve this multi-objective optimization problem, a multi-objective quantum-inspired cultural algorithm is proposed, which adopts the dual structure to effectively utilize the implicit knowledge extracted from the non-dominating individuals set to promote more efficient search. It has three highlights. One is the rectangle’s height of each allele is calculated by non-dominated sort among individuals. The second is the crowding degree that records the density of non-dominated individuals in the topological cell measure the uniformity of the Pareto-optimal set instead of the crowding distance. The third is the update operation of quantum individuals and the selection operator are directed by the knowledge. Simulation results indicate that the layout of wireless sensor network obtained by this algorithm have larger network cover rate and less node redundancy rate.

Yinan Guo, Dandan Liu, Meirong Chen, Yun Liu

Artificial Bee Colony Algorithm for Modular Neural Network

The Artificial bee colony (ABC) algorithm is simple, robust and has been used in the optimization of synaptic weights from an Artificial Neural Network (ANN). However, this is not enough to generate a robust ANN. Modular neural networks (MNNs) are especially efficient for certain classes of regression and classification problems, as compared to the conventional monolithic artificial neural networks. In this paper, we present a model of MNN based on ABC algorithm (ABC-MNN). Experiments show that, compared to the monolithic ABC-NN model, classifier designed in this model has higher training accuracy and generalization performance.

Chen Zhuo-Ming, Wang Yun-Xia, Ling Wei-Xin, Xing Zhen, Xiao Han-Lin-Wei

An Intelligent Optimization Algorithm for Power Control in Wireless Communication Systems

High instantaneous peak power of the transmitted signals is the main obstacle of orthogonal frequency division multiplexing (OFDM) systems for its application, therefore, the peak to average power ratio (PAPR) reduction has been one of the most important technologies. Among all the existing methods, partial transmit sequences (PTS) is a distortionless phase optimization technique that significantly improves PAPR performance to with a small amount of redundancy. However, the computational complexity in conventional PTS increases exponentially with the number of subblocks. In this paper, an intelligent optimization method is proposed for PTS technique to obtain good balance between computational complexity and PAPR performance. Simulation results show that the proposed method can achieve better performance compared with conventional algorithms.

Jing Gao, Jinkuan Wang, Bin Wang, Xin Song

Optimized Neural Network Ensemble by Combination of Particle Swarm Optimization and Differential Evolution

The Neural-Network Ensemble (NNE) is a very effective method where the outputs of separately trained neural networks are combined to perform the prediction. In this paper, we introduce the improved Neural Network Ensemble (INNE) in which each component forward neural network (FNN) is optimized by particle swarm optimization (PSO) and back-propagation (BP) algorithm. At the same time, the ensemble weights are trained by Particle Swarm Optimization and Differential Evolution cooperative algorithm(PSO-DE). We take two obviously different populations to construct our algorithm, in which one population is trained by PSO and the other is trained by DE. In addition, we incorporate the fitness value from last iteration into the velocity updating to enhance the global searching ability. Our experiments demonstrate that the improved NNE is superior to existing popular NNE.

Zeng-Shun Zhao, Xiang Feng, Fang Wei, Shi-Ku Wang, Mao-Yong Cao, Zeng-Guang Hou

Feature Analysis, Clustering, Pattern Recognition and Classification

SOR Based Fuzzy K-Means Clustering Algorithm for Classification of Remotely Sensed Images

Fuzzy k-means clustering algorithms have successfully been applied to digital image segmentations and classifications as an improvement of the conventional k-means cluster algorithm. The limitation of the Fuzzy k-means algorithm is its large computation cost. In this paper, we propose a Successive Over-Relaxation (SOR) based fuzzy k-means algorithm in order to accelerate the convergence of the algorithm. The SOR is a variant of the Gauss–Seidel method for solving a linear system of equations, resulting in faster convergence. The proposed method has been applied to classification of remotely sensed images. Experimental results show that the proposed SOR based fuzzy k-means algorithm can improve convergence speed significantly and yields comparable similar classification results with conventional fuzzy k-means algorithm.

Dong-jun Xin, Yen-Wei Chen

UMPCA Based Feature Extraction for ECG

In this paper, we propose an algorithm for 12-leads ECG signals feature extraction by Uncorrelated Multilinear Principal Component Analysis(UMPCA). However, traditional algorithms usually base on 2-leads ECG signals and do not efficiently work out for 12-leads signals. Our algorithm aims at the natural 12-leads ECG signals. We firstly do the Short Time Fourier Transformation(STFT) on the raw ECG data and obtain 3rd-order tensors in the spatial-spectral-temporal domain, then take UMPCA to find a Tensor-to-Vector Projection(TVP) for feature extraction. Finally the Support Vector Machine(SVM) classifier is applied to achieve a high accuracy with these features.

Dong Li, Kai Huang, Hanlin Zhang, Liqing Zhang

Genetic Algorithm Based Neural Network for License Plate Recognition

This paper combines genetic algorithms and neural networks to recognize vehicle license plate characters. We train the neural networks using a genetic algorithm to find optimal weights and thresholds. The traditional genetic algorithm is improved by using a real number encoding method to enhance the networks weight and threshold accuracy. At the same time, we use a variety of crossover operations in parallel, which broadens the range of the species and helps the search for the global optimal solution. An adaptive mutation rate both ensures the diversity of the species and makes the algorithm convergence more rapidly to the global optimum. Experiments show that this method greatly improves learning efficiency and convergence speed.

Wang Xiaobin, Li Hao, Wu Lijuan, Hong Qu

Optimizing Fuzzy ARTMAP Ensembles Using Hierarchical Parallel Genetic Algorithms and Negative Correlation

This study demonstrates a system and methods for optimizing a pattern classification task. A genetic algorithm method was employed to optimize a Fuzzy ARTMAP pattern classification task, followed by another genetic algorithm to assemble an ensemble of classifiers. Two parallel tracks were performed in order to assess a diversity-enhanced classifier and ensemble optimization methodology in comparison with a more straightforward method that does not rely on diverse classifiers and ensembles. Ensembles designed with diverse classifiers outperformed diversity-neutral classifiers in 62.50% of the tested cases. Using a negative correlation method to manipulate inter-classifier diversity, diverse ensembles performed better than non-diverse ensembles in 81.25% of the tested cases.

Chu Kiong Loo, Wei Shiung Liew, Einly Lim

Improvement of Panchromatic IKONOS Image Classification Based on Structural Neural Network

Remote sensing image classification plays an important role in urban studies. In this paper, a method based on structural neural network for panchromatic image classification in urban area with adaptive processing of data structures is presented. Backpropagation Through Structure (BPTS) algorithm is adopted in the neural network that enables the classification more reliable. With wavelet decomposition, an object’s features in wavelet domain can be extracted. Therefore, the pixel’s spectral intensity and its wavelet features are combined as feature sets that are used as attributes for the neural network. Then, an object’s content can be represented by a tree structure and the nodes of the tree can be represented by the attributes. 2510 pixels for four classes, road, building, grass and water body, are selected for training a neural network. 19498 pixels are selected for testing. The four categories can be perfectly classified using the training data. The classification rate based on testing data reaches 99.91%. In order to prove the efficiency of the proposed method, experiments based on conventional method, maximum likelihood classification, are implemented as well. Experimental results show the proposed approach is much more effective and reliable.

Weibao Zou

Local Feature Coding for Action Recognition Using RGB-D Camera

In this paper, we perform activity recognition using an inexpensive RGBD sensor (Microsoft Kinect). The main contribution of this paper is that the conventional STIPs feature are extracted from not only the RGB image, but also the depth image. To the best knowledge of the authors, there is no work on extracting STIPs feature from the depth image. In addition, the extracted feature are combined under the framework of locality-constrained linear coding framework and the resulting algorithm achieves better results than state-of-the-art on public dataset.

Mingyi Yuan, Huaping Liu, Fuchun Sun

Circular Projection for Pattern Recognition

There are a number of methods that transform 2-D shapes into periodic 1-D signals so that faster recognition can be achieved. However, none of these methods are both noise-robust and scale invariant. In this paper, we propose a circular projection method for transforming 2-D shapes into periodic 1-D signals. We then apply a number of feature extraction methods to the 1-D signals. Our method is invariant to the translation, rotation and scaling of the 2-D shapes. Also, our method is robust to Gaussian white noise. In addition, it performs very well in terms of classification rates for a well-known shape dataset.

Guangyi Chen, Tien Dai Bui, Sridhar Krishnan, Shuling Dai

A Tensor Factorization Based Least Squares Support Tensor Machine for Classification

In the fields of machine learning, image processing, and pattern recognition, the existing least squares support tensor machine for tensor classification involves a non-convex optimization problem and needs to be solved by the iterative technique. Obviously, it is very time-consuming and may suffer from local minima. In order to overcome these two shortcomings, in this paper, we present a tensor factorization based least squares support tensor machine (TFLS-STM) for tensor classification. In TFLS-STM, we combine the merits of least squares support vector machine (LS-SVM) and tensor rank-one decomposition. Theoretically, TFLS-STM is an extension of the linear LS-SVM to tensor patterns. When the input patterns are vectors, TFLS-STM degenerates into the standard linear LS-SVM. A set of experiments is conducted on six second-order face recognition datasets to illustrate the performance of TFLS-STM. The experimental results show that compared with the alternating projection LS-STM (APLS-STM) and LS-SVM, the training speed of TFLS-STM is faster than those of APLS-STM and LS-SVM. In term of testing accuracy, TFLS-STM is comparable with LS-SVM and is superiors to APLS-STM.

Xiaowei Yang, Bingqian Chen, Jian Chen

A Remote Sensing Image Classification Method Based on Extreme Learning Machine Ensemble

There are few training samples in the remote sensing image classification. Therefore, it is a highly challenging problem that finds a good classification method which could achieve high accuracy and strong generalization to deal with those data. In this paper, we propose a new remote sensing image classification method based on extreme learning machine (ELM) ensemble. In order to promote the diversity within the ensemble, we do feature segmentation and nonnegative matrix factorization (NMF) to the original data firstly. Then ELM is chosen as base classifier to improve the classification efficiency. The experimental results show that the proposed algorithm not only has high classification accuracy, but also handles the adverse impact of few training samples in the classification of remote sensing well both on the remote sensing image and UCI data.

Min Han, Ben Liu

Model Identification of an Unmanned Helicopter Using ELSSVM

The dynamic model of unmanned helicopter is a coupled nonlinear system. With respect to the identification problem for this model, extended least squares support vector machine (ELSSVM) is proposed. ELSSVM extends the solution space of structure parameters to improve the convergence performance. Base width of kernel function and regularization parameter of ELSSVM are minimized by differential evolution (DE). As compared to the traditional identification method for helicopter dynamic model, the proposed method omits the linear process and the trained model is closer to the helicopter dynamic model. The data-driven based experiments show that the proposed method takes a short training time and has a high identification accuracy.

Xinjiu Mei, Yi Feng

A Feature Point Clustering Algorithm Based on GG-RNN

In the field of object recognition in computer vision, feature point clustering algorithm has become an important part of the object recognition. After getting the object feature points, we make the feature points in clustering in the use of GG-RNN clustering algorithm, to achieve multi-part of the object clustering or the multi-object clustering. And the GG-RNN clustering algorithm we propose innovatively, is merged with the grayscale and gradient information based on Euclidean distance in the similarity calculation. Compared with the distance description of basic RNN algorithm, the similarity calculation of high-dimensional description of GG-RNN will improve the accuracy of the clustering in different conditions.

Zhiheng Zhou, Dongkai Shen, Lei Kang, Jie Wang

Local Fisher Discriminant Analysis with Locally Linear Embedding Affinity Matrix

Fisher Discriminant Analysis (FDA) is a popular method for dimensionality reduction. Local Fisher Discriminant Analysis (LFDA) is an improvement of FDA, which can preserve the local structures of the feature space in multi-class cases. However, the affinity matrix in LFDA cannot reflect the actual interrelationship among all the neighbors for each sample point. In this paper, we propose a new LFDA approach with the affinity matrix being solved by the locally linear embedding (LLE) method to preserve the particular local structures of the specific feature space. Moreover, for nonlinear cases, we extend this new LFDA method to the kernelized version by using the kernel trick. It is demonstrated by the experiments on five real-world datasets that our proposed LFDA methods with LLE affinity matrix are applicable and effective.

Yue Zhao, Jinwen Ma

A Facial Expression Recognition Method by Fusing Multiple Sparse Representation Based Classifiers

We develop a new method to recognize facial expressions. Sparse representation based classification (SRC) is used as the classifier in this method, because of its robustness to occlusion. Histograms of Oriented Gradient (HOG) descriptors and Local Binary Patterns are used to extract features. Since the results of HOG+SRC and LBP+SRC are complimentary, we use a classifier combination strategy to fuse these two results. Experiments on Cohn-Kanade database show that the proposed method gives better performance than existing methods such as Eigen+SRC, LBP+SRC and so on. Furthermore, the proposed method is robust to assigned occlusion.

Yan Ouyang, Nong Sang

Image Data Classification Using Fuzzy c-Means Algorithm with Different Distance Measures

Fuzzy c-Means algorithms(FCMs) with different distance measures are applied to an image classification problem in this paper. The distance measures discussed in this paper are the Euclidean distance measure and divergence distance measure. Different distance measures yield different types of Fuzzy c-Means algorithms. Experiments and results on a set of satellite image data demonstrate that the classification model employing the divergence distance measure can archive improvements in terms of classification accuracy over the models using the FCM and SOM algorithms which utilize the Euclidean distance measure.

Dong-Chul Park

Global Matching to Enhance the Strength of Local Intensity Order Pattern Feature Descriptor

Local intensity order pattern feature descriptor is proposed to extract the feature of image recently. However, it did not provide the global information of an image. In this paper, a simple, efficient and robust feature descriptor is presented, which is realized by adding the global information to local intensity features. A descriptor, which utilizes local intensity order pattern and/or global matching, is proposed to gather the global information with local intensity order. Experimental results shows that the proposed hybrid approach outperform over the state-of-the art feature extraction method like scale-invariant feature transform, local intensity order pattern and DAISY for standard oxford dataset.

Hassan Dawood, Hussain Dawood, Ping Guo

An Approach of Power Quality Disturbances Recognition Based on EEMD and Probabilistic Neural Network

Based on intrinsic mode functions (IMFs), standard energy difference of each IMF obtained by EEMD and probabilistic neural network (PNN), a new method is proposed to the recognition of power quality transient disturbances. In this method, ensemble empirical mode decomposition (EEMD) is used to decompose the non-stationary power quality disturbances into a number of IMFs. Then the standard energy differences of each IMF are used as feature vectors. At last, power quality disturbances are identified and classified with PNN. The experimental results show that the proposed method can effectively realize feature extraction and classification of single and mixed power quality disturbances.

Ling Zhu, Zhigang Liu, Qiaoge Zhang, Qiaolin Hu

Chinese Text Classification Based on Neural Network

Text classification is widely used nowadays. In this paper, we proposed a combination feature reduction method to reduce feature space dimension based on inductive analysis of existing researches. Neural network was then trained and used to classify new documents. Existing researches mainly focus on the classification of the English text, but we focused on classification of Chinese text instead in this paper. Experimental results showed that the proposed feature reduction method performed well, and the neural network needed less terms to achieve the same accuracy compared with other classifiers.

Hu Li, Peng Zou, WeiHong Han

The Angular Integral of the Radon Transform (aniRT) as a Feature Vector in Categorization of Visual Objects

The recently introduced angular integral of the Radon transform (aniRT) seems to be a good candidate as a feature vector used in categorization of visual objects in a rotation invariant fashion. We investigate application of aniRT in situations when the number of objects is significant, for example, Chinese characters. Typically, the aniRT feature vector spans the diagonal of the visual object. We show that a subset of the full aniRT vector delivers a good categorization results in a timely manner.

Andrew P. Papliński

A Fast Algorithm for Clustering with MapReduce

MapReduce is a popular model in which the dataflow takes the form of a directed acyclic graph of operators. But it lacks built-in support for iterative programs, which arise naturally in many clustering applications. Based on micro-cluster and equivalence relation, we design a clustering algorithm which can be easily parallelized in MapReduce and done in quite a few MapReduce rounds. Experiments show that our algorithm not only runs fast and obtains good accuracy but also scales well and possesses high speedup.

Yuqing Miao, Jinxing Zhang, Hao Feng, Liangpei Qiu, Yimin Wen

Gaussian Message Propagation in d-order Neighborhood for Gaussian Graphical Model

Gaussian graphical models are important undirected graphical models with multivariate Gaussian distribution. A key probabilistic inference problem for the model is to compute the marginals. Exact inference algorithms have cubic computational complexity, which is intolerable for large-scale models. Most of approximate inference algorithms have a form of message iterations, and their computational complexity is closely related to the convergence and convergence rate, which causes the uncertain computational efficiency. In this paper, we design a fixed parameter linear time approximate algorithm — the Gaussian message propagation in

d

-order neighborhood. First, we define the

d

-order neighborhood concept to describe the propagation scope of exact Gaussian messages. Then we design the algorithm of Gaussian message propagation in

d

-order neighborhood, which propagates Gaussian messages in variable’s

d

-order neighborhood exactly, and in the (

d

 + 1)th-order neighborhood partly to preserve the spread of the Gaussian messages, and computes the approximate marginals in linear time

O

(

n

·

d

2

) with the fixed parameter

d

. Finally, we present verification experiments and comparison experiments, and analyze the experiment results.

Yarui Chen, Congcong Xiong, Hailin Xie

Fast Image Classification Algorithms Based on Random Weights Networks

Up to now, rich and varied information, such as networks, multimedia information, especially images and visual information, has become an important part of information retrieval, in which video and image information has been an important basis. In recent years, an effective learning algorithm for standard feed-forward neural networks (FNNs), which can be used classifier and called random weights networks (RWN), has been extensively studied. This paper addresses the image classification algorithms using the algorithm. A new algorithm of image classification based on the RWN and principle component analysis (PCA) is proposed. The proposed algorithm includes significant improvements in classification rate, and the extensive experiments are performed using challenging databases. Compared with some traditional approaches, the new method has superior performances on both classification rate and running time.

Feilong Cao, Jianwei Zhao, Bo Liu

A Convolutional Neural Network for Pedestrian Gender Recognition

We propose a discriminatively-trained convolutional neural network for gender classification of pedestrians. Convolutional neural networks are hierarchical, multilayered neural networks which integrate feature extraction and classification in a single framework. Using a relatively straightforward architecture and minimal preprocessing of the images, we achieved 80.4% accuracy on a dataset containing full body images of pedestrians in both front and rear views. The performance is comparable to the state-of-the-art obtained by previous methods without relying on using hand-engineered feature extractors.

Choon-Boon Ng, Yong-Haur Tay, Bok-Min Goi

The Novel Seeding-Based Semi-supervised Fuzzy Clustering Algorithm Inspired by Diffusion Processes

Semi-supervised clustering can take advantage of some labeled data called seeds to bring a great benefit to the clustering of unlabeled data. This paper uses the seeding-based semi-supervised idea for a fuzzy clustering method inspired by diffusion processes, which has been presented recently. To investigate the effectiveness of our approach, experiments are done on three UCI real data sets. Experimental results show that the proposed algorithm can improve the clustering performance significantly compared to other semi-supervised clustering approaches.

Lei Gu

Fault Detection for Nonlinear Discrete-Time Systems via Deterministic Learning

This paper presents a fault detection scheme for nonlinear discrete-time systems based on the recently proposed deterministic learning (DL) theory. The scheme consists of two phases: the learning phase and the detecting phase. In the learning phase, the discrete-time system dynamics underlying normal and fault modes are locally accurately approximated through deterministic learning. The obtained knowledge of system dynamics is stored in constant RBF networks. In the detecting phase, a bank of estimators are constructed using the constant RBF networks to represent the learned normal and fault modes. By comparing the set of estimators with the monitored system, a set of residuals are generated, and the average

L

1

norms of the residuals are used to compare the differences between the dynamics of the monitored system and the dynamics of the learning normal and fault modes. The occurrence of a fault can be rapidly detected in a discrete-time setting.

Junmin Hu, Cong Wang, Xunde Dong

Loose Particle Classification Using a New Wavelet Fisher Discriminant Method

Loose particles left inside aerospace components or equipment can cause catastrophic failure in aerospace industry. It is vital to identify the material type of these loose particles and eliminate them. This is a classification problem, and autoregressive (AR) model and Learning Vector Quantization (LVQ) networks have been used to classify loose particles inside components. More recently, the test objects have been changed from components to aerospace equipments. To improve classification accuracy, more data samples often have to be dealt with. The difficulty is that these data samples contain redundant information, and the aforementioned two conventional methods are unable to process redundant information, thus the classification accuracy is deteriorated. In this paper, the wavelet Fisher discriminant is investigated for loose particle classifications. First, the fisher model is formulated as a least squares problem with linear-in-the-parameters structure. Then, the previously proposed two-stage subset selection method is used to build a sparse wavelet Fisher model in order to reduce redundant information. Experimental results show the wavelet Fisher classification method can perform better than AR model and LVQ networks.

Long Zhang, Kang Li, Shujuan Wang, Guofu Zhai, Shaoyuan Li

L 1 Graph Based on Sparse Coding for Feature Selection

In machine learning and pattern recognition, feature selection has been a very active topic in the literature. Unsupervised feature selection is challenging due to the lack of label which would supply the categorical information. How to define an appropriate metric is the key for feature selection. In this paper, we propose a “filter” method for unsupervised feature selection, which is based on the geometry properties of ℓ

1

graph. ℓ

1

graph is constructed through sparse coding. The graph establishes the relations of feature subspaces and the quality of features is evaluated by features’ local preserving ability. We compare our method with classic unsupervised feature selection methods (Laplacian score and Pearson correlation) and supervised method (Fisher score) on benchmark data sets. The classification results based on support vector machine, k-nearest neighbors and multi-layer feed-forward networks demonstrate the efficiency and effectiveness of our method.

Jin Xu, Guang Yang, Hong Man, Haibo He

Vision Modeling and Image Processing

An Image Segmentation Method for Maize Disease Based on IGA-PCNN

The image segmentation of plant diseases is one of the critical technical aspects of digital image processing technology for Disease Recognition. This paper proposes an improved pulse coupled neural network based on an improved genetic algorithm. An objective evaluation function is defined based on linear weighted function with maximum Shannon entropy and minimum cross-entropy. Through adaptive adjustment of crossover probability and mutation probability, we optimized the parameters of pulse coupled neural network based on the improve genetic algorithm. The improved network is used to segment the color images of Maize melanoma powder disease in RGB color subspaces. Then combined with the results by color image merger strategy, we can get the terminal results of target area. The experimental results show that this method could segment the disease regions better and set complexity parameters simplier.

Wen Changji, Yu Helong

A Vector Quantization Approach for Image Segmentation Based on SOM Neural Network

In the existing segmentation algorithms, most of them take single pixel as processing unit and segment an image mainly based on the gray value information of the image pixels. However, the spatially structural information between pixels provides even more important information of the image. In order to effectively exploit both the gray value and the spatial information of pixels, this paper proposes an image segmentation method based on Vector Quantization (VQ) technique. In the method, the image to be segmented is divided into small sub-blocks with each sub-block constituting a feature vector. Further, the vectors are classified through vector quantization. In addition, the self-organizing map (SOM) neural network is proposed for realizing the VQ algorithm adaptively. Simulation experiments and comparison studies have been conducted with applications to medical image processing in the paper, and the results validate the effectiveness of the proposed method.

Ailing De, Chengan Guo

A Fast Approximate Sparse Coding Networks and Application to Image Denoising

Sparse modeling has proven to be an effective and powerful tool that leads to state of the art algorithms in image denoising, inpainting, super-resolution reconstruction, etc. Although various sparse modeling algorithms have been proposed, a major problem of these algorithms is computationally expensive which prohibits them from real-time applications. In this paper, we propose a simple and efficient approach to learn fast approximate sparse coding networks as well as show its application to image denoising. Our experiments demonstrate that the pre-learned network is over 200 times faster than sparse optimization algorithm, and yet obtain approving result in image denoising.

Jianyong Cui, Jinqing Qi, Dan Li

Residual Image Compensations for Enhancement of High-Frequency Components in Face Hallucination

Recently a growing interest has been seen in single-frame super-resolution techniques, which are known as example-based or learning based super-resolution techniques. Face Hallucination is one of such techniques, which is focused on resolution enhancement of facial images. Though face hallucination is a powerful and useful technique, some detailed high-frequency components cannot be recovered. In this paper, we propose a high-frequency compensation framework based on residual images for face hallucination method in order to improve the reconstruction performance. The basic idea of proposed framework is to reconstruct or estimate a residual image, which can be used to compensate the high-frequency components of the reconstructed high-resolution image. Three approaches based on our proposed framework are proposed. Experimental results show that the high-resolution images obtained using our proposed approaches can improve the quality of those obtained by conventional face hallucination method.

Yen-Wei Chen, So Sasatani, Xianhua Han

Recognition Approach of Human Motion with Micro-accelerometer Based on PCA-BP Neural Network Algorithm

A human motion recognition method based on micro-acceleration sensor technology is put forward in this paper. Acceleration information acquire system is designed, which is including a tri-axial accelerometer, a micro-processor, a wireless transmission module and power supply program. The signal preprocessing and methods of feature extraction is analyzed. What’s more, the experiment of human hand motion recognition based on BP neural network is carried out, results show that method proposed have recognition rate of 90%, compare the characteristics without processing and through principal component analysis (PCA) respectively after the identification experiment, the results show that the latter improve recognition effect and speed up convergence rate.

Yuxiang Zhang, Huacheng Li, Shiyi Chen, Liuyi Ma

A Stable Dual Purpose Adaptive Algorithm for Subspace Tracking on Noncompact Stiefel Manifold

Starting from an extended Rayleigh quotient defined on the noncompact Stiefel manifold, in this paper, we present a novel dual purpose subspace flows for subspace tracking. The proposed algorithm can switch from principal subspace to minor subspace tracking with a simple sign change of its stepsize parameter. More interestingly, the proposed dual purpose gradient system behaves the same invariant property as that of the well-known Chen-Amari-Lin system. The stability of the discrete version of the proposed subspace flow is guaranteed by an additional added stabilizing term. No tunable parameter is required for the proposed algorithm as opposed to the modified Oja algorithm. The strengths of the proposed algorithm is demonstrated using a

de

facto

benchmark example.

Lijun Liu, Yi Xu, Qiang Liu

Invariant Object Recognition Using Radon and Fourier Transforms

In this paper, an invariant algorithm for object recognition is proposed by using the Radon and Fourier transforms. It has been shown that this algorithm is invariant to the translation and rotation of pattern images. The scaling invariance can be achieved by the standard normalization techniques. Our algorithm works even when the center of the pattern object is not aligned well. This advantage is because the Fourier spectra are invariant to spatial shift in the radial direction whereas existing methods assume the centroids are aligned exactly. Experimental results show that the proposed method is better than the Zernike’s moments, the dual-tree complex wavelet (DTCWT) moments, and the auto-correlation wavelet moments for one aircraft database and one shape database.

Guangyi Chen, Tien Dai Bui, Adam Krzyzak, Yongjia Zhao

An Effective Method for Signal Extraction from Residual Image, with Application to Denoising Algorithms

To minimize image blurring and detail loss caused by denoising, we propose a novel method to exploit residual image. Firstly, we apply Non-local Means (NLM) filter to original image to get the denoised image and store the weights used for averaging. Secondly, we filter the residual image with the stored weights. Then a Gaussian filter is applied to the denoised residual image before we add the results to image denoised by NLM to recover the lost image details. Different from previous methods, our method uses the structure information in the original image and can be used to extract lost image details from residual images with very low SNR. An analysis on the mechanism of the signal extraction method is given. Quantitative evaluation showed that the proposed algorithm effectively improved accuracy of NLM filter. In addition, the residual of the final results contained fewer observable structures, demonstrating the effectiveness of the proposed method to recover lost details.

Min-Xiong Zhou, Xu Yan, Hai-Bin Xie, Hui Zheng, Guang Yang

Visual Attention Computational Model Using Gabor Decomposition and 2D Entropy

Visual attention is an important mechanism as it can be applied to many branches of computer vision and image processing such as segmentation, compression, detection, tracking and so on. Based on both capabilities and defects of existing models, the paper proposes a computational saliency-oriented model from the perspective of frequency domain. A saliency map can be generated by two main steps: firstly Gabor wavelet decomposition of the input image at certain levels is used to produce the feature components, and then these components are selected and fused in the sense of 2D entropy. The proposed algorithm outperforms most of state-of-the-art algorithms at human fixation prediction for both psychological patterns and natural images including salient objects with arbitrary sizes. Beyond that, biological plausibility of Gabor filter makes our approach more reliable and adaptive to various stimuli.

Qi Lv, Bin Wang, Liming Zhang

Human Detection Algorithm Based on Bispectrum Analysis for IR-UWB Radar

Impulse-radio Ultra-wide band (IR-UWB) radar plays an important role in searching and detecting human target in particular situations, such as counterterrorism, post-disaster search and rescue and so on. It mainly takes the advantages of its good penetrability through obstacles and high range resolution. It detects human target mainly by detecting the respiratory signal. As the higher order spectrum is immune to the Gaussian noise, a new algorithm based on the bispectrum analysis for human detection behind the wall is proposed. The results of the through-wall experiments show the algorithm has a better performance than the conventional PSD-based algorithm.

Miao Liu, Sheng Li, Hao Lv, Ying Tian, Guohua Lu, Yang Zhang, Zhao Li, Wenzhe Li, Xijing Jing, Jianqi Wang

Backmatter

Weitere Informationen