Lecture Notes in Electrical Engineering Volume 56 Qi Luo (Ed.) Advancing Computing, Communication, Control and Management ABC QiLuo SchoolofElectricalEngineering WuhanInstituteofTechnology Wuhan430070 China E-mail:[email protected] ISBN978-3-642-05172-2 e-ISBN978-3-642-05173-9 DOI10.1007/978-3-642-05173-9 LectureNotesinElectricalEngineering ISSN1876-1100 LibraryofCongressControlNumber:2009940122 (cid:2)c 2010Springer-VerlagBerlinHeidelberg Thisworkissubject tocopyright. Allrightsarereserved, whetherthewholeorpart ofthematerial isconcerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation,broad- casting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of thispublicationorpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyrightLaw ofSeptember 9, 1965, initscurrent version, andpermission for use must always be obtained from Springer.ViolationsareliabletoprosecutionundertheGermanCopyrightLaw. Theuseofgeneraldescriptivenames,registerednames,trademarks,etc.inthispublicationdoesnot imply, even in the absence of a specific statement, that such names are exempt from the relevant protectivelawsandregulationsandthereforefreeforgeneraluse. Typeset&CoverDesign:ScientificPublishingServicesPvt.Ltd.,Chennai,India. Printedinacid-freepaper 987654321 springer.com Preface A large 2008 ISECS International Colloquium on Computing, Communication, Control, and Management (CCCM 2008), was held in Guangzhou, August 2008, China. Just like the name of the Colloquium, the theme for this conference is Advancing Computing, Communication, Control, and Management Technologies. 2008 ISECS International Colloquium on Computing, Communication, Control, and Management is co-sponsored by Guangdong University of Business Studies, China, Peoples’ Friendship University of Russia, Russia, Central South University, China, Southwestern University of Finance & Economics, China, and University of Amsterdam, Netherlands. It is also co-sponsored IEEE Technology Management Council, IEEE Computer Society, and Intelligent Information Technology Application Research Institute. Much work went into preparing a program of high quality. We received about 972 submissions. Every paper was reviewed by 3 program committee members, about 382 were selected as regular papers, representing a 39% acceptance rate for regular papers. The CCCM conferences serve as good platforms for the engineering community to meet with each other and to exchange ideas. The conference has also stroke a balance between theoretical and application development. The conference committees have been formed with over two hundred committee members who are mainly research center heads, faculty deans, department heads, professors, and research scientists from over 30 countries. The conferences are truly international meetings with a high level of participation from many countries. The response that we have received for the congress is excellent. This volume contains revised and extended research articles written by prominent researchers participating in the conference. Topics covered include intelligent computing, network management, wireless networks, telecommunication, power engineering, control engineering, Signal and Image Processing, Machine Learning, Control Systems and Applications, The book will offer the states of arts of tremendous advances in Computing, Communication, Control, and Management and also serve as an excellent reference work for researchers and graduate students working on Computing, Communication, Control, and Management Research. Qi Luo Table of Contents Study on MRF and REF to Semi-supervised Classification ............ 1 Liang Jun, Xianyi Cheng, and Xiaobo Chen An Extension Method of Space Syntax and Application............... 7 Xinqi Zheng, Lu Zhao, Yanjun Su, Guang Yan, and Shuqing Wang A Feasible Registration Method for Underwater SLAM ............... 15 Feng Sun, Wenjing Wang, Fuqiang Liu, and Wenfeng Wang A Promoted Global Convergence Particle Swarm Optimization Algorithm....................................................... 23 Du Ronghua and Cai Yue A Web-Based Integrated System for Construction Project Cost Prediction ...................................................... 31 Huawang Shi and Wanqing Li Research of Corporate Credit for Anhui Province’s Listed Companies Based on Computer Technology.................................... 39 Li Yang and Malin Song Evaluation of Industrial Parks’ Industrial Transformations and Environmental Reform Actualized by AHP Based on MatLab Software ........................................................ 47 Malin Song A Model Integrated Development of Embedded Software for Manufacturing Equipment Control ................................. 55 Zhumei Song and Di Li Two Stage Estimation Method in Data Processing and Simulation Analysis ........................................................ 62 Pan Xiong, Chunru Zhao, and Lihong Jin Research and Implementation of a Reconfigurable Parallel Low Power E0 Algorithm.................................................... 71 Li Wei, Dai Zibin, Nan Longmei, and Zhang Xueying A Model of Car Rear-End Warning by Means of MAS and Behavior.... 79 Liang Jun, Xianyi Cheng, Xiaobo Chen, and Yao Ming Novel Hue Preserving Algorithm of Color Image Enhancement......... 88 Caixia Liu VIII Table of Contents Artificial Immune System Clustering Algorithm and Electricity Customer Credit Analysis......................................... 97 Shu-xia Yang Virtual Space Sound Technique and Its Multimedia Education Application ..................................................... 103 Xinyu Duan Survey on Association Rules Mining Algorithms ..................... 111 Mingzhu Zhang and Changzheng He Application of Rough Set Theory and Fuzzy LS-SVM in Building Cooling Load.................................................... 119 Xuemei Li, Ming Shao, and Lixing Ding An Efficient Feature Selection Algorithm Based on Hybrid Clonal Selection Genetic Strategy for Text Categorization ................... 127 Jiansheng Jiang, Wanneng Shu, and Huixia Jin Power Demand Forecasting Based on BP Neural Network Optimized by Clone Selection Particle Swarm ................................. 135 Xiang Li and Shi-jun Lu Research on Simplifying the Motion Equations and Coefficients Identification for Submarine Training Simulator Based on Sensitivity Index........................................................... 142 Zhao Lin and Zhu Yi Evaluating Quality of Networked Education via Learning Action Analysis ........................................................ 150 Bin Xu Research on Face Recognition Technology Based on Average Gray Scale ........................................................... 158 Weihua Wang The Active Leveled Interest Management in Distributed Virtual Environment .................................................... 166 Jia Bei and Yang Zhao The Research of the Intelligent Fault Diagnosis Optimized by ACA for Marine Diesel Engine............................................. 174 Peng Li, Lei Liu, and Haixia Gong Extraction and Parameterizationof Eye Contour from Monkey Face in Monocular Image ................................................ 182 Dengyi Zhang, Chengzhang Qu, Jianhui Zhao, Zhong Zhang, Youwang Ke, Shizhong Han, Mingqi Qiao, and Huiyun Zhang Table of Contents IX KPCA and LS-SVM Prediction Model for Hydrogen Gas Concentration ................................................... 190 Minqiang Pan, Dehuai Zeng, Gang Xu, and Tao Wang Random Number Generator Based on Hopfield Neural Network and SHA-2 (512) .................................................... 198 Yuhua Wang, Guoyin Wang, and Huanguo Zhang A Hybrid Inspection Method for Surface Defect Classification ......... 206 Mengxin Li, Chengdong Wu, and Yang Cao Study on Row Scan Line Based Edge Tracing Technology for Vehicle Recognition System .............................................. 214 Weihua Wang Study on Modeling and Simulation of Container Terminal Logistics System ......................................................... 222 Li Li and Wang Xiaodong Digitalized Contour Line Scanning for Laser Rapid Prototyping........ 231 Zeng Feng, Yao Shan, and Ye Changke 3D Surface Texture Synthesis Using Wavelet Coefficient Fitting........ 239 Muwei Jian, Ningbo Hao, Junyu Dong, and Rong Jiang BuildingServiceOrientedSharingPlatformforEmergencyManagement – An Earthquake Damage Assessment Example...................... 247 Ying Su, Zhanming Jin, and Jie Peng An Interactive Intelligent Analysis System in Criminal Investigation.... 256 Ping He Research on Functional Modules of Gene Regulatory Network ......... 264 Honglin Xu and Shitong Wang Block-BasedNormalized-Cut Algorithm for Image Segmentation ....... 272 Haiyu Song, Xiongfei Li, Pengjie Wang, and Jingrun Chen Semantics Web Service Characteristic Composition ApproachBased on Particle Swarm Optimization...................................... 279 Zhou Xiangbing Author Index.................................................. 289 Study on MRF and REF to Semi-supervised Classification Liang Jun, Xianyi Cheng, and Xiaobo Chen School of Computer Science and Telecommunication Engineering, Jiangsu University, China, 212013 [email protected] Abstract. We study the semi-supervised classifier with a decision rule learning from labeled and unlabeled data. A model to semi-supervised classification is proposed to overcome the problem induced by mislabeled samples. A new energy function based on REF (robust error function) is used in MRF (Markov Random Field). Also two algorithms based on iterative condition mode and Markov chain Monte Carlo respectively are designed to infer the label of both labeled and unlabeled samples. Our experiments demonstrate that the proposed method is efficient for real-world dataset. Keywords: semi-supervised learning, classifier, Markov Random Field, Simulating. 1 Introduction Semi-supervised learning has received considerable attention in the machine learning literature due to its potential in reducing the need for expensive labeled data [1]. Given a sample setY ={y ,y ,...,y ,y ,...,y }⊂ Rmand a label setL={1,2,......,C}, the 1 2 l l+1 N first l samples y(1≤i≤l) are labeled as f ∈L and the remaining i i samplesy(l+2≤i≤N) are unlabeled. The goal is to classify the unlabeled samples to its i latent class. Most formulations of semi-supervised learning approach the problem from one of the two ends of the unsupervised-supervised spectrum: either supervised learning in the presence of unlabelled data [2,3] or unsupervised learning with additional information [4,5]. Almost all current research has not considered the situation when there exist mislabeled samples in semi-supervised classification. Because of the little number of labeled samples in semi-supervised classification, the mislabeled samples will have severe influence on the final classification result. An example is show in figure 1 which contains 70 samples from two classes. One class is show in blue and the other in red. There are eight labeled samples in the data set which are show in green. Three of them are from class 1, four of them are from class 2 and especially one sample from class 1 is mislabeled to class 2 (show by the green diamond on top most). Q. Luo (Ed.): Advancing Computing, Communi., Control and Management, LNEE 56, pp. 1–6. springerlink.com © Springer-Verlag Berlin Heidelberg 2010 2 L. Jun, X. Cheng, and X. Chen According to the algorithm in literature [6], we can see that many samples in class 1 are misclassified to class 2 because of the existence of mislabeled sample. In order to tackle the problem, we proposed a new model based on MRF and REF. The paper is organized as follows. In section 2, we proposed a MRF with REF. In Sect 3, we describe two different inference algorithm for the model and Section 4 reports the experiments on the above examples and its application in semi-supervised classification. Some conclusions are given in section 5. 2 MRF with REF 2.1 MRF for Semi-supervised Classification We first recall MRF (MRF) models [7] which originated in statistical physics and have been extensively used in image processing. MRF constraints the spatial smoothness though puts high probability to the class labels which are consistent locally. In the context of machine learning, what we can do is create a graph with a node for each sample, and with undirected edges between them that are similar to each other. A typical MRF model is show in figure 1. The nodes in the upper level are that we can observe while the nodes in the lower level are unobserved. Because of its flexibility, MRF have been introduced into semi-supervised learning [8].In the context of semi-supervised classification, some hidden labels are known as prior which differentiate it from the conventional MRF as illustrated in figure 2. A natural energy function takes the form: E (x)=−∑ωx x g ij i j (1) {i,j} wherex ∈{−1,+1} are binary labels and ωis the weight on edge{i, j}, which is a i ij measure of the similarity between the samples. 2.2 REF Recently, many researchers have found that the use of REF can lead to improved results on applications such as super-resolution, CCD demosaicing, image in-painting, and image denoising. Examples of REFs include the truncated quadratic function and the Lorentzian error function [9] which is 1 E (x)=ln(1+ x2) (2) l 2 The most important feature of REF is that unlike the quadratic error function, the magnitude of the derivative of the error function does not increase as the error rises. This feature is show in figure 3 which we can see the rate of increase in the Lorentzian error function actually decreases as the error rises. It is this property that makes this function robust to large outlier errors. Study on MRF and REF to Semi-supervised Classification 3 2 2 1.5 1.5 1 1 0.5 0.5 0 0 -0.5 -0.5 -1 -1 -1.5 -1.5 -20 0.5 1 1.5 -20 0.5 1 1.5 Fig. 1. Classification results according to the semi-supervised method of [6] 7 6 5 4 3 2 1 -030 -20 -10 0 10 20 30 Fig. 2. MRF Fig. 3. Lorentzian error function 2.3 MRF with REF Due to the analysis above, we incorporate the global energy function of MRF which means a good classifying function should not change too much between nearby samples with the local Lorentzian error function which means a good classifying function should not change too much from the initial label assignment into the follow energy function: 1 E(x)=λE(x)+(1−λ)E(x)=λ∑ω(x−x)2+(1−λ)∑log(1+ (x−f)2) (3) g l ij i j 2 i i {i,j} 1≤i≤l whereω =exp(−||y −y ||2 /2σ2), which is a measures of the similarity between ij i j λ samples. is a parameter which balance the global consistency and local consistency. To assign a probability distribution to the labels, we form a random field: 1 p(x)= exp{−βE(x)} (4) Z where the partition functionZ normalizes over all labels. So our goal is to maximizep(x)which is equivalent to minimizeE(x)as below: x=argminE(x) (5) x