Foundations of Intelligent Systems

In this paper, a novel algorithm which is called CIPMPBE (Closed Iterative Pattern Miner via Prime-Block Encoding) is proposed to mine closed iterative patterns. CIPMPBE is composed of three separated steps. In the first step, the positional information of all frequent

i

-sequences is generated. In the second step, the positional information of all frequent

i-

sequences and all instances of frequent (

i-

1)-iterative patterns are used to obtain the positional information of all instances of frequent

i

-iterative patterns. In the third step, the positional information of all instances of frequent

i

-iterative pattern is used to obtain the positional information of the entire closed

i

-iterative pattern and get back to the first step to generate the positional information of closed

(i+

1)-iterative pattern. For effective testing, a set of experiments were performed. The results of these experiments show that the time efficiency of CIPMPBE is better than that of CLIPER (CLosed Iterative Pattern minER).

Active learning is an important method for solving data scarcity problem in machine learning, and most research work of active learning are pool-based. However, this type of active learning is easily affected by pool size, and makes performance improvement of classifier slow. A novel active learning with constructing queries based pool is proposed. Each iteration the training process first chooses representative instance from pool predefined, then employs climbing algorithm to construct instance to label which best represents the original unlabeled set. It makes each queried instance more representative than any instance in the pool. Compared with the original pool based method and a state-of-the-art active learning with constructing queries directly, the new method makes the prediction error rate of classifier drop more fast, and improves the performance of active learning classifier.

This paper studies the modeling method of soft measurement based on the permeability of index sintering process. Soft sensor modeling can use three kinds of methods, neural network, fuzzy control and adaptive fuzzy neural network of soft measurement based on subtraction clustering. Through analysis various soft sensor modeling methods and experimental data, Subtraction clustering adaptive fuzzy neural network method has a very good convergence, and also it has highly precision of prediction, smaller test error, so the method very suitable for the engineering application.

Cancer classification of gene expression data helps determine appropriate treatment and the prognosis. Accurate prediction to the type or size of tumors relies on adopting efficient classification models such that patients can be provided with better treatment to therapy. In order to gain better classification, in this study, a linear relevant feature dimensionality reduction method termed the neighborhood preserving based semi-supervised dimensionality reduction (NPSSDR) is applied. Different from traditional supervised or unsupervised methods, NPSSDR makes full use of side information, which not only preserves the must-link and cannot-link constraints but also can preserve the local structure of the input data in the low dimensional embedding subspace. Experimental results using public gene expression data show the superior performance of the method.

Service robot should have human emotions, and farthest resemble the real-life interaction when interact with humans. A layered model of artificial emotion called AME (Attitude Mood Emotion) model, merging with attitude is proposed, which is more human, and an emotion-stimulation generation mechanism based on the OCC cognitive appraisal model and the PAD emotion space is built. The concept of “Mood Baseline” is introduced to describe the interaction among emotions, mood and attitude. The validity and practicability of this model can be verified based on “FuNiu” robot.

As a universally accepted tool of machine learning, support vector machine (SVM) is efficent in most scenarios but often suffers from prohibitive complexity in dealing with large-scale classification problems in terms of computation time and storage space. To address such intractability, this paper presents a group and nearest neighbor strategy aiming to extract support vectors from training samples for obtaining the discriminant function in a fast fashion as only the support vectors contribute to the function. For non-linear cases, kernel function is investigated and adopted in this approach. The proposed stragtegy is described through mathematical analysis and evaluated by a set of numerical experiments. The result demonstrates that the suggested approach is effective in addressing the large-scale classification problems with acceptabe complexity.

A local feature saliency approach is introduced which locally selects relevant features for particular clusters. In addition, the original feature saliency algorithm is improved to deal with redundant features by using local correlations. Experiments using 10 benchmark data sets are conducted for estimating the performances of proposed approach. For comparison, results obtained from the traditional algorithms are also provided. The experimental results show that the proposed local feature selection approach outperforms the traditional algorithms.

This paper gives an approach for fusing the directed acyclic graphs (DAGs) of BNs, an important and popular probabilistic graphical model (PGM). Considering conditional independence as the semantics implied in a BN, we focus on the DAG fusion while preserving the semantics in all participating BNs. Based on the concept and properties of Markov equivalence, we respectively give the algorithms for fusing equivalent and inequivalent common subgraphs of all participating BNs.

In this study, we consider the problems associated with selecting and scheduling a set of R&D projects to maximize the overall net present value. This paper proposes a zero-one integer programming model in conjunction with a genetic algorithm (GA) to overcome these problems. Taguchi Method was employed in the design of the GA parameters to increase the efficiency of the proposed method. We conclude that the proposed GA is capable of efficiently solving problems associated with the management of portfolios.

The author has been engaging in the research of Tibetan language information processing techniques, and has presided over the development of CDMA 450M network-based Tibetan mobile phone and car phone[1],[5], as well as smart mobile phone Tibetan software package based on Windows Mobile and Android. In 2010, required by China Mobile Communications Group Co., Ltd. Tibet Branch, Symbian S60 v3 operating system based smart mobile phones was developed. Through in-depth study of Symbian operating system, the author firstly proposed Tibetan processing key technology implemented on Symbian S60 v3 based smart mobile phone.

Pulse Coupled Neural Network (PCNN) model has been widely used in digital image processing, but its parameters determination is always under a difficult situation. This paper analyzed the firing time of the coupled linking PCNN, and indicates the difference between theoretical firing time and actual firing time when the neuron is under the influence of the neighboring neurons. By analyzing the influence of the parameters on the coupling effects of neighboring neurons, a new method of setting parameters is proposed in image segmentation. Revealing that only with the proper parameters setting, can the theoretical firing time and the actual firing time of the neuron be consistent, so that the pulse burst characteristics of PCNN can be truly realized. For Lena image segmentation, etc, the similar effects is obtained with the traditional experience parameters, and showing validity and efficiency of our proposed method.

The classic algorithms of Hierarchical Task Network (HTN) planning focus on searching valid solutions with knowledge learning or heuristic mechanisms. However, there is little work on the planning optimization, especially on handling the problem in the situation that there is little heuristic knowledge to guide the optimum solution search process while many non-optimum solutions exist, or there is significance interactions between the abstract intermediate goals as which are not independent. This paper put forward a messy genetic algorithm (MGA) to solve the optimum solution searching problem of HTN planning in the above situations. Length-variant chromosome is introduced to represents the possible planning solution in form of decomposition tree with dynamic node numbers. Simulation results indicate that the MGA can locate the optimum solution among the huge search space with about 3×2

14

possible solutions within 6 seconds.

This paper studies semantic retrieval method based on rule-based reasoning under support of ontology. In the support of ontology knowledge base, this system faces ontology structure-based metadata, according to related concepts and reasoning rules, looks for implication relationship and digs out implicit information. We extract reasoning rules from semantic relationship of audit information, and present Smallest Connected Subgraph Generation Algorithm, Connected Componets Generation Algorithm and Connected Componets Matching Rules Algorithm; these three algorithms are in solving problems in the process of reasoning, implement the reasoning function. Through case analysis, we verify the application probability of these algorithms in audit information resource retrieval.

Multiple approaches aim at describing numerical data with words. In this paper we give a brief overview of a distributed agent-based system providing summaries of city traffic in a textual form. The system deals locally with a problem of incomplete data adopting a method for grounding of modal statements in artificial cognitive agents. The internal uncertainty of an agent about a current state of the traffic is expressed with autoepistemic modal operators of possibility, belief, and knowledge.

This paper research Paris-MOU NIR (New Inspection Regime) of latest PSC rules of 2009/16/EC, Combining the rough set theory with hierarchic analysis model in view of not reduce efficient classify information, introducing the definition of importance into non-core attributes, a new data reduction algorithm is proposed and put forward an algorithm of PSC ship-selecting system. The results reveal that the algorithm is able to resolve data reduction problem and simplify network structure.

Artificial immune algorithm and ant colony algorithm are combined to deal with problem of 2D route planning of aircraft. Initial routes are generated randomly within the flying area and clonal selection algorithm is used to search good routes. A group of routes with minimum cost of threat and oil are gained. Some initial pheromone is put nearby these routes. Based on this, ant colony algorithm are used to search optimal route while threaten avoid and minimum cost are taken into consideration.

Support vector machines (SVM) together with principal component analysis (PCA) have been applied to hyperspectral image classification and mapping with great success. PCA has been proved to be an effective preprocessing tool for dimension reduction and/or feature extraction. After dimension reduction with PCA, the classification and mapping time can be dramatically reduced while retaining good accuracy. However, the computational cost of PCA preprocessing can be as high as that of SVM classification applied to the original unreduced data set. Researchers have studied different algorithms to cut the PCA preprocessing time, while some others totally ignored that cost. We propose a simple PCA preprocessing scheme which can reduce the computational complexity many folds and is particularly suitable for image data of large size. High classification accuracy can be achieved. A numerical example on an Earth Observing-1 (EO-1) Hyperion image is included to demonstrate the viability of this new procedure. In addition, this example clearly shows that the standard PCA preprocessing may require as much time as the SVM classification does.

Rough set theory is a mathematic tool to handlevague and uncertain knowledge. The core problem of rough set theory is the reduction of attributes. This paperintroduces apresentation of the information granularity for ordered information systems based on dominance relation, and detail the way to calculate the attribute importance based on the information granularity we have defined. On that basis, put forward a heuristic attribute reduction algorithm for ordered information systems. Experiments show that the algorithm performs well and outperforms some other rivals.

Strategy designing is a key task in today’s market place competition and many other application fields. While dominating over competitors is unrealistic or too expensive to seek after, finding dynamic equilibrium in the competition progress becomes critical and practical in lots of scenarios. Inspired by the Abrams and Strogatz model originally proposed for monitoring language death and competition, this paper presented a novel mechanism for agent’s competition equilibrium point finding and strategy designing. By analyzing historical data and status property of the two competition agents, an agent could evaluate its own status and forecast the long range trends of the game based upon the improved Abrams and Strogatz model. A posteriori status evaluation based strategy designing protocol is purposed and implemented. Experiments and simulation results showed desirable consistence with theoretical analysis. The protocol presented is valuable for adoption with proper settings to practical applications.

On the basis of

α

-quasi-lock semantic resolution method in lattice-valued propositional logic (

$\mathcal{L}_{n}\times\mathcal{L}_{2}$

)P(X),

α

-quasi-lock semantic resolution in linguistic truth-valued lattice-valued propositional logic

$\mathcal{L}_{V(n\times2)}$

P(X) is studied in the present paper. Firstly, (

c

i

,

t

)-quasi-lock semantic resolution for

$\mathcal{L}_{V(n\times2)}$

P(X) is equivalently transformed into that for lattice-valued propositional logic

$\mathcal{L}_{Vn}$

P(X). Secondly, similar equivalence between (

c

i

,

f

)-quasi-lock semantic resolution for

$\mathcal{L}_{V(n\times2)}$

P(X) and that for

$\mathcal{L}_{Vn}$

P(X) is also established under certain conditions.

Recently more and more people pay attention to software GPS receiver, because it provides a high level of flexibility and simplicity for many potential applications. The resulting software GPS receivers offer considerable flexibility in modifying settings to accommodate new applications without redesigning hardware, choosing an IF frequency, or implementing future upgrades. This paper introduces the design of a software-based receiver for L1-band civilian GPS applications, including the design of the RF front end based on MAX2769, the algorithms for the signal acquisition, tracking and positioning decoding, which are demonstrated.

According to the general form of principle of duality in the sense of class [1], this paper tries to study the dual operators of operators in lattice implication algebra [2], especially the dual operator of implication operator and gives the expression of principle of duality in lattice implication algebra. Then we will show its value in developing theories in corresponding field.

Kernel construction is one of the key issues both in current research and application of kernel methods. In this paper, we present an effective kernel construction method, in which we reduce the construction of kernel function to the solutions of generalized eigenvalue problems. Specifically, we first obtain a primal kernel function based on the similarity of instances, and refine it with the conformal transformation. Then we determine the parameters of kernel function by simply solving the generalized eigenvalue problems according to the kernel alignment and Fisher criteria. Our method can avoid local maxima and insure positive semidefinite of the constructed kernel. Experimental results show that our kernel construction method is effective and robust.

Model combination is an important approach to improving the generalization performance of support vector machine (SVM), but usually has low computational efficiency. In this paper, we propose a novel probabilistic model combination method for support vector machine on regularization path (PMCRP). We first design an efficient regularization path algorithm, namely the regularization path of support vector machine based on positive-definite kernel (PDSVMP), which constructs the initial candidate model set. Then, we combine the initial models using Bayesian model averaging. Experimental results on benchmark datasets show that PMCRP has significant advantage over cross-validation and the Generalized Approximate Cross-Validation (GACV), meanwhile guaranteeing high computation efficiency of model combination.

In this paper, we focus on the properties of filters in lattice implication algebra. We study the relationship of associative filter and implicative filter, n-fold associative filter and n-fold implicative filter in detail. And a sufficient condition of involution filter in lattice implication algebra is proved. Then the relationship of some filters is given in Figure 4, and the transformation conditions among these filters are analyzed and obtained in Figure 5. Last, some properties of filter lattice are discussed.

We introduce the heterogeneous vision radius which follows powerlaw distribution into the evolutionary prisoner’s dilemma game among mobile agents to study the evolution of cooperation. It is found that the cooperation level of the system presents remarkable differences when vision radius follows power-law distribution with different exponents. The cooperative agents can survive in such a system provided that the heterogeneity of vision radius is moderate and the temptation of defection is not too high. This work may be beneficial to understand cooperative behavior in biological and social systems consisted of mobile agents.

In order to communicate successfully highly distributed agents must utilise a shared naming convention. Such a naming convention can be developed by the agents through the process of semiosis where the agents collectively establish the naming convention for the objects. Narrowing the interaction pattern of Language Game Model to a single speaker (teacher) we model the process of individual semiosis. Further, using a developed simulation framework of the process, we analyse the dynamic behaviour of the alignment against deviations from the idealised settings of shared context. In particular, we study the reaction on the formation of naming convention among the interacting population in settings where the communication channels are subject to imperfection. As such, this research fills the current gap and investigates the dynamics of the fine-grained model of semiosis. In particular, we analyse both, analytically and through simulation, the observed phase transition in the alignment process.

For many applications over the Web of Linked Data, it is a key challenge to access multiple distributed data sources in an integrated and transparent way. In traditional federated query, data from unknown data sources are ignored, which in turn leads to a poor recall of the query result. Furthermore, due to the openness of the Web of Linked Data, it is very difficult to know in advance all data sources in this Web. To overcome these problems we present a novel approach to query the Web of Linked Data. The main idea of our approach is to discover query services that might be used to answer a query during the query execution itself. Our approach can be independently used or the complement of traditional federated query for querying this Web of Data. We provide a prototype implementation and performance evaluation of our work. The experiment shows the feasibility of our approach.

In this paper, a data imputation method with a Support Vector Machine (SVM) is proposed to solve the issue of missing data in activity-based diaries. Here two SVM models are established to predict the missing elements of ‘number of cars’ and ‘driver license’. The inputs of the former SVM model include five variables (Household composition, household income, Age oldest household member, Children age class and Number of household members). The inputs of the latter SVM model include three variables (personal age, work status and gender). The SVM models to predict the ‘number of cars’ and ‘driver license’ can achieve accuracies of 69% and 83% respectively. The initial experimental results show that missing elements of observed activity diaries can be accurately inferred by relating different pieces of information. Therefore, the proposed SVM data imputation method serves as an effective data imputation method in the case of missing information.

The automatic identification of security vulnerabilities in the binary code is still a young but important research area for the security researchers. In recent years, the number of identified integer overflow vulnerabilities has been increasing rapidly. In this paper, we present a smart software vulnerability detection technology, which is used for the identification of integer overflow vulnerabilities in the binary executables. The proposed algorithm is combined with debugger module, static analysis module and genetic algorithm module. We use the fitness function to guide the generation of the tested data and use static analysis to provide the information that the genetic module needs. Theory analyses and experiment results indicate that the detection technology based upon genetic algorithm can identify the exceptions in the object program and is more efficient than the common Fuzzing technology.

The circle and rectangle packing problem with equilibrium constraints is difficult to solve in polynomial time. In this paper a case based reasoning genetic algorithm is presented to improve its solving efficiency and accuracy. Its main idea is that the initial population of genetic algorithm consists of the reasoning solutions attained from the case-solutions and its non feasible layout solutions generated randomly; in each iteration step of the evolutionary process, a non-isomorphic layout pattern solution of the current optimal individual is constructed automatically and replaces the worse one in the population. The experimental results demonstrate that the proposed algorithm has higher computational speed and accuracy than HCIGA.

Feature ordering is important in Incremental Attribute Learning where features are gradually trained in one or more size. Apart from time-consuming contribution-based feature ordering methods, feature ordering also can be derived by filter criteria. In this paper, a novel criterion based on a new metric called Discriminability is presented to give ranks for feature ordering. Final results show that the new metric not only is applicable for IAL, but also exhibits better performance in lower error rates.

In some research situations, we often have to classify data with incomplete values which affect the learning performance of classifiers. Although various classification algorithms have been proposed, most of them are short of the ability to deal with incomplete data. This paper proposes a novel approach based on selective ensemble for classifying incomplete data. The method finds the local complete patterns for which the feature values are complete and trains multiple component learners for each local complete subset. Then, it combines the outputs of the classifiers. The method needs no assumption about the incomplete mechanism that is necessary for previous methods. The proposed method is evaluated by three datasets from the UCI Machine Learning Repository. The experiments results show that classification accuracy of the proposed method is superior to those of widely used imputations and deletion method.

Dealing with vague or imprecise information has been always a challenging problem. Different tools have been proposed to manage that uncertainty. A new model based on hesitant fuzzy sets was presented to manage situations where experts hesitate among several values to assess alternatives, variables, etc. Hesitant fuzzy sets models quantitative settings, however, it could occur similar situations but in qualitative settings, where experts think of several possible linguistic values or richer expressions than a single linguistic term to assess alternatives, variables, etc. In this contribution the aim is to introduce the concept of

Hesitant Fuzzy Linguistic Term Sets

(HFLTS) that will provide a linguistic elicitation based on the fuzzy linguistic approach and the use of context-free grammars.

Situation Awareness is the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning and the projection of their status in the near future. It is a crucial factor in decision-making in a dynamic environment particularly with certain degrees of risk, called risk-based situation awareness. In this paper we first explore the most popular models in situation awareness, data fusion and risk assessment. We show how they complement each other in developing a conceptual model for risk-based situation awareness. We will also demonstrate how this model can be used to support decision-making in a dynamic environment.

A general overview of partial comparability and preference theory allows examining the notion of bipolarity and its role in the development of some general preference structures. This bipolar approach comes natural to the framework of decision theory, where different preference structures can be initially explored according to the type of bipolar model that they follow. Therefore, we compare two general preference structures, the first one, referred to as the PCT structure, which results from a well known axiomatic model for partial comparability theory, and the second one, referred to as the P-A structure, which extends one particular standard fuzzy preference model, such that some basic differences as well as particular similarities are clearly identified.

In this paper, a novel honesty-rate measuring based approach is proposed to improve the security and trust of distributed intrusion detection systems. All the cooperative nodes join the system with an initial value of 1 for an honesty rate. The honesty rate of a node dynamically varies depending on its status and the honesty rate will decrease when a suspicious behavior of the node is detected and increase when the node obeys certain criteria. The proposed approach compares the honesty rate of each node to eliminate or reduce the impact of harmful information from the malicious nodes, and then reduces the false positives and false negatives of the intrusion detection systems. The experiments and analyses of a representative case confirm the ability of the proposed approach to improve detection accuracy and detection capability.

In this paper, we present a multiclass classification algorithm to address the multiclass problems with cooperative clustering. Using cooperative clustering, the cluster centers of all classes can be computed iteratively and simultaneously. In the process of clustering, we select a pair of adjacent class, and make their cluster center drawn towards the boundary. Therefore, the data set around a class is found and the data set plus the data in this class can be trained to form a classifier. With this algorithm, training efficiency and classification efficiency are improved with a slight impact on classification accuracy.

Let

G

be a planar graph with maximum degree ∆ ≥ 7 and without 3-cycles adjacent to 4-cycles, that is, any 3-cycle has not a common edge with a 5-cycle. Then the total chromatic number of

G

is ∆ + 1.

This paper presents a novel approach for managing multi-bilateral concurrent negotiations. We extend our previous work by considering a situation where a buyer agent negotiates with multiple seller agents concurrently over multiple continuous issues instead of a single issue. A related work in this area considers a meta-strategy for bilateral negotiations. This work adapts the previous related work to coordinate multi-bilateral concurrent negotiations taking into consideration the different behaviors of the opponents during negotiation to decide on choosing the appropriate negotiation strategy (i.e., trade-off or concession) for the buyer agent’s delegates at each negotiation round. A negotiation meta-strategy to coordinate the one-to-many negotiation form is proposed and empirically tested under various negotiation environments. The experiments show the robustness of our coordination mechanism.

In this paper we combining LHIA research the generated filters of lattice implication algebra. Firstly some new properties of lattice implication algebra are discussed. Then a structure of generated filter of two filters’ union is obtained. Finally some elementary properties are given and a simple proof of(

F

(

L

), ∨ , ∧ )is a complete distributive lattice is also obtained.

Under complex, dynamic and uncertain environment, tasks in multi-agent system need to be distributed to agents dynamically and agents need to cooperate to complete tasks assigned dynamically. This paper proposes a dynamic task allocation model based on game theory and dynamic coordination in task execution based on coordination graph at each time step. The synthesized model is solved by reinforcement learning. The detailed algorithm is illustrated with an example and experimental results show that the method is an effective solution for dynamic task allocation and execution coordination under uncertain environment.

This paper was carried out mainly for the problem of learner model in adaptive learning system, such as the unscientific attention dimension, poor calculated representation method and single and subjective method to obtain feature values. We have put forward a new learner model to achieve self-organization of learning resources on the basis of learning goals and learner’s personal conditions. It includes three feature items such as knowledge level, cognitive ability and preferences. We respectively introduce the representation and acquisition of feature value of learner model in detail. After that, we propose a push mechanism of learning resources. Experimental results show this learner model is effective and practical in the application.

Most of the Information Retrieval (IR) systems are built on the Client – Server paradigm. While agent-based market systems have relative success, agent-based IR systems seem to fail. Recently, works on a mobile agent-based approach for the search of information on the World Wide Web arise. Indeed, the emergence of mobile agent has given the researchers a new way to achieve efficient mobile agents-based IR systems. This paradigm certainly holds great promise, though the lack of results in reliability and security issues. However, we feel that integrating market mechanisms to a new mobile agent model should improve security while bringing the possibility to solve non market applications problems. In this paper, we present how to use a mobile agent-based approach. The main idea is to generalize the market mechanisms to non-market systems such as IR through an extended mobile agents’ model, the seller – buyer model. Then, we present the architecture of a secure mobile agent-based searching system that derives from a general mobile agents-based architecture. Finally, we give an experimental validation to our proposition.

Agent-based negotiation support has been applied in electronic commerce and provides users with valuable information about products and services. In this paper a framework for goal-driven negotiation process is proposed which can be used to design negotiating agents. The framework includes goal representation, goal selection, goal execution and goal reconsideration. Goals are set to control agents’ inference processes and determine negotiation actions. Instead of focusing on predicting possible negotiation outcomes, the framework emphasizes rational and explainable deliberation process of negotiating agents.

Multi-agent systems are distributed problem solving systems involving multiple collaborating intelligent agents that are capable of interacting with their environments. Toward the goal of developing multi-agent systems of bounded rationality, an important issue is how to manage and handle inconsistent or conflicting knowledge and information an agent may possess or has to reason with. The focus of this paper is on inconsistency in multi-agent systems. We describe the occurrence of inconsistency in the depth of knowledge, define nine different types of inconsistent phenomena, and discuss possible ways to utilize inconsistency as useful heuristics toward developing multi-agent systems of bounded rationality. The main contribution is that we shed some new light on inconsistency in multi-agent systems.

With a very broad application prospect, emotional speech recognition has aroused many researchers attentions. The key to the problem is how to extract the features which can express feelings exactly. In this paper, the extraction method based on speech syllable distribution is presented. In contrast with the traditional one the new method can reduce the statistical error without computing the speech pause, moreover, the features such as syllable time and mute time are introduced to aiding emotional speech recognition.

Since wavelet transform can not fully describe facial curves features, in this paper, we propose a novel face recognition method based on Curvelet domain and kernel principal component analysis (KPCA). Using multi-scale, multi-directional Curvelet transform to extract image features not only has higher approximation accuracy and better performance of sparse expression, but also can effectively express the singularity along the curve. Furthermore, kernel principal component analysis (KPCA) is used to project Curvelet feature coefficient into kernel space with more expressing capability. Finally, the nearest method is adopted to classify. The results indicate that the algorithm is effective in image dimension reduction and face recognition rate in the JAFFE face database, ORL face database and FERET face database.

To solve the problem that the accuracy of SAR image segmentation is not high enough in the marine spill oil detection, a segmentation method of marine spill oil images based on maximum entropy and CV model is proposed in this paper. Firstly, the multilevel threshoding algorithm based on maximum entropy is used to make a coarse segmentation for marine spill oil images. The obtained spill oil region and coarse contour provide local region and initial contour for CV model, respectively, to reduce the scene complexity of CV model and its sensitivity to initial situation. That is CV model is utilized to subdivide the local area. Lots of experimental results show that the proposed segmentation method of marine spill oil SAR images not only enables the dispense with initial condition but also ensures accurate segmentation contour and efficient operation.

The method of threshold selection based on maximal Shannon entropy or exponent entropy only depend on the probability information from gray image histogram, and don’t immediately consider the uniformity of the gray scale within the cluster. Considering these facts, thresholding based on exponent gray entropy is proposed. Firstly, exponent gray entropy is defined and the method of single threshold selection is given. Then, the method is extended to multilevel thresholding. Furthermore, the niche chaotic mutation particle swarm optimization algorithm is adopted to find the best multi-threshold. Many experimental results show that, compared with multilevel thresholding based on maximal entropy and particle swarm optimization, the proposed segmentation method has less operation times and segmented images of the suggested method are more accurate in edge and texture.

Representing images using Bag-of-Words (BOW) model has been shown excellent performance for image classification and retrieval. However, there are still some limitations in this model such as the presence of many noisy visual words and the hard to define vocabulary size. To circumvent these drawbacks, this paper concentrates on tuning compact, robust and thus efficient BOW model even with a universal size for image representation. The proposed approach increases expressive power by employing Sparse Partial Least Squares (SPLS) for tuning the traditional and high-dimensional BOW model and learning more discriminative subspace with 10 latent variables. The performance of learning BOW models to image classification is studied through extensive experiments on the VOC 2006 dataset. Empirical results indicate that the proposed method yields quite stable results, and outperforms the traditional BOW models with various vocabulary sizes and PCA with SVM.

In accordance with the structural features of QR barcode, a text blind watermarking algorithm is proposed in this paper. The algorithm is that the Hill

2

Encryption and BCH error correction encoding are used in the watermark pro-processing and then the watermark information is embedded in the rows and columns of QR barcode by using structural fine-tuning method. The QR barcode security and confidentiality are effectively improved by combining two-dimensional barcode with digital watermark technology.

This study addresses a cloth pattern simulation method on the basis of a 1/

f

noise, which is used to generate 3D controlled curves for the objective cloth. Users can conveniently get the different cloth pattern effects only by controlling the key points’ positions of the 1/f noise function. The curved surface can be fitted by these 3D controlled curves. For the purpose of enhancing the visualization, a texture mapping method is employed and some improvements are made upon the illumination model. Finally, the experimental results show that the method is efficient and has potential research importance in the cloth texture simulation.

Since most images coming from nature show the fractal characteristic, the fractal dimension can be used to quantitative characterization and analysis on these images. However, noise may have an effect on the fractal dimension. In this work, the effects of salt & pepper noise, Gaussian white noise and multiplicative noise on the fractal dimension are studied by using different material images. The study shows that the three kinds of noises all have a significant effect on the fractal dimension. Digital images with added noise become coarser, and their corresponding fractal dimensions increase. An anti-noise determination on fractal dimension based on differential box counting (DBC) algorithm and noise character is proposed. Pixel values in a box are all used and the deviation between the maximum and minimum within the box is replaced with double standard deviation. Compared with the general pretreatment method, the proposed method is valid and convenient.

The matching and retrieval of shapes is an important challenge in computer vision. A large number of shape similarity approaches have been developed. In this paper, we employ two approaches for improving shape retrieval. First, we use angle gradient to extract contour’s critical points, this is a simple approach which decrease the computational cost while retain spatial information of shapes. Second, we present a pairwise similarity measure, which is a quadratic assignment problem. This problem is approximately solved by spectral technique. This method is tested on standard MPEG-7 shape database using the standard performance evaluation scheme. The experimental results indicate that the proposed method outperforms the closely relate method.

In this paper, a skew detection of fabric images scheme based on morphological method and projection profile analysis is proposed. The original image is convoluted using Sobel operator, and the weft boundary is generated by binarizing the convoluted image according to setting optimal threshold. A projection profile is computed at each angle, and the angle that maximizes a criterion function is regarded as the skew angle. Experimental results show that the skew angles detected by the proposed detection algorithm with higher accuracy.

An improved method of separation and identification of touching kernels and foreign materials in digital images is proposed. The touching kernels are separated by using watershed algorithm based on morphological multiscale decomposition (MSD). Then, feature extraction of kernels is used for calculation of Mahalanobis distance. Finally foreign materials are identified by comparing Mahalanobis distance with the given threshold. The performance of the new algorithm is compared to that of a method based on a watershed algorithm and Mahalanobis distance (WMD). The experimental results showed that the efficiency of the proposed algorithm was superior with regard to WMD.

In this paper, we present a novel object detection scheme that uses information of the sample fragments. These sample fragments are extracted by decomposition of the sample contour. Then, the candidate fragments corresponding to the sample fragments are detected from the images by partial Hausdorff distance. The Multiclass Discriminative Field (MDF) is used to select the most probable fragments from candidate fragments. The parameter estimation and inference of the MDF are simplified by using the candidate fragments as nodes of a graph. With these selected fragments, the contours of the objects can be obtained. The experiments on our postmark database and the ETHZ database show the feasibility of our proposed scheme.

The characteristics of the Daubechies wavelet are studied. How to realize the Mallat algorithm and how to calculate the denoising threshold are discussed. A group of new application components on wavelet denoising are developed on JBeam platform. Finally, the components are utilized to denoise the real signal of bench test of automotive with sample cases, and the denoising effect of each Daubechies mother wavelet is analyzed to prove that the most suitable mother wavelet is Db14.

In order to quickly get the optimal threshold for image segmentation, a novel method of image segmentation based on improved adaptive genetic algorithm is presented, which adjusts the parameters of genetic algorithm adaptively according to the difference among the individuals and the diversity of population to ensure the convergence of the algorithm and avoid the precocious. The method combines with the Otsu method to get the best image segmentation threshold. The theoretically analysis and simulation experiments show that the threshold is more accurate and less time is consumed greatly by using the proposed method than Otsu image segmentation and other adaptive genetic algorithm.

In order to promote the accuracy and practicability of resistance identifier feature recognition in the defects inspection of PCB (Printed Circuit Board), a resistance texture feature description and recognition method combining wavelet transform and Ojala’s LBP (Local Binary Pattern) operator is proposed in this paper. Firstly, wavelet analysis is adopted to decompose the original resistance image for dimension reduction. Then the approximate image is divided into several sub-blocks, and two types of sub-block LBP histogram is drawn by using two different LBP operators. Finally, we concatenated the whole histogram of every sub-blocks into histogram sequence, and the sequence is the enhanced feature vector of resistance image recognition. Experimental results show that the proposed method has a high-level recognition rate of texture feature.

Fast and reliable image processing method is one of the key technologies for realizing welding seam tracking. To this end, a new image processing method for V-shape weld seam based on laser structured light is proposed. This method starts with image preprocessing, including median filtering, one-dimensional LOG (Laplace of Gaussian) operations, morphological operations, and scattered clusters filtering. For improving the efficiency of image processing, it can find a target window dynamically. Then, the local search method is used twice for improving the accuracy of extracting the laser centerline. At last, this laser centerline is fitted by the modified least square method, and the feature points of V-shape weld seam can be obtained by calculating the intersections of the fitted lines. The results of image processing show that this method can quickly and reliably process seam images with strong interference, and meet the requirements of welding seam tracking.

A new LCV (local chan-vess) model algorithm based on prior shape focusing on segmentation of melon image is proposed in this paper to measure the micro changes of morphological parameters. During the process, local boundary information image of melon is acquired firstly through mathematical morphology algorithm, and construct LCV model according to the form of traditional CV model. After that, a prior shape can be obtained by mathematical morphology and spline interpolation algorithms, and then be integrated into LCV model functional through a shape comparing function, thus the LCV model based on prior shape is constructed. Compared with traditional edge detection and segmentation algorithms, the new algorithm proposed in this paper could obtain more ideal boundary information.

In the exploration of the planets of our solar system, images taken during a lander’s descent to the surface of a planet provide a critical link for surface reconstruction. In use of descent images taken by landing camera is an effective way for terrain mapping. The method by dividing virtual planes for reconstruction has been improved in our paper. Based on plane homograph relationship, the match-related value noise is eliminated by gauss filter, and discontinuity disparity areas are smoothed by introduction of smooth-constraint match cost function. We demonstrate experimental results on synthetic images and our method can achieve better accuracy and effectiveness for terrain mapping.

In connection with user perception-related features in the best view of 3D object, a method of best view selection of 3D object based on sample learning is proposed. Firstly, the AdaBoost algorithm is applied on the sample set for supervised learning in order to get the best view space of 3D object. Then, the rendering views come from different viewpoints in 3D object are compared and analyzed with the best view in shape similarity matching. The rendering view with highest similarity will be regarded as the best view of 3D object. The experiments show that the best view selected is very consistent with the user’s subjective choices. This method has good validity and reliability.

An improved retina modeling based on bilateral filter is developed for face recognition under variable lighting. Bilateral filter is applied to estimate the adaptation factor X0 (local lighting) in the Naka-Rushton equation, which models nonlinear processing of photoreceptors and outer plexiform layers. Difference of Gaussians filter (DoG) is also used to enhance image contours so as to model inner plexiform layer. Experimental results on the Yale B and CMU PIE face databases indicate the effectiveness of the proposed method.

This work describes the development of an automated method to measure the length of filament-shaped objects from Atomic Force Microscopy (AFM), Scanning Tunneling Microscopy (STM) and Transmition Electron Microscopy (TEM) images. The proposed methodology can determine the length of the object(s) of interest using image segmentation, where the Parametric Deformable Model (PDM) (especially the Snake Model - SM) was applied. The measurement procedure starts from the segmentation step, where an improved erosion is applied, resulting in a thin curve representation with only two endpoints (skeleton). The piecewise linear approximation concept was also employed to measure the total length of an object, based on the dimension of the pixel (unitary length) that composes the skeleton. The accuracy of the algorithm was evaluated using a high precision STM 7x7 silicon image and a sort of DNA filaments (AFM).

Leukocyte differential count is a standard process in hematological diagnosing. The abnormal leukocytes impede the development of robust and efficient computer-assisted blood film analysis system, especially the smudge cells which are highly variable due to scrunching. This paper concentrates on the feature space construction for smudge cells. Unimodal regional features and weak edge characters are combined with histogram distribution statistics and gray level co-occurrence matrix to form the final feature space. A Supporting Vector Machine is adopted to construct the classifier. Experiment shows that the proposed features are efficient to compute and the sensitivity and specificity for smudge cell classification are both promising.

The presentation of touching erythrocytes affects the Red Blood Cell counting and diagnosing in automatic blood cell analysis based on digital image processing, thus a robust and efficient recognition algorithm for touching erythrocytes is required. This paper analyses the typical features of touching erythrocytes and presents two new features to characterize them. The algorithm quantifies the radial uniformity of cell outer contours to describe the shape of the erythrocyte, and calculates the contour distribution within sector regions to efficiently approximate the gray scale distribution for texture representation. An ANN classifier was constructed based on the above features, and the discriminating rate for touching erythrocytes achieved 91.82%, which is promising and indicates the efficiency of the proposed features.

Recently, there are increasing attention paid on compressed sensing which is distinct different from traditional signal processing and image reconstructed from indirect or incomplete measurement, especially ℓ 1-norm problem which is transformed from compressive sensing. The idea of ℓ 1_regularization, as the one of ℓ 1-norm, has been receiving a lot of interest in signal processing, image recovery and statistic, etc. This paper will introduces a continuation log-barrier method for solving ℓ 1_regularized least squares problem in the field of compressive sensing, which is a second-order method. Our work is inspired by the work in [4] and continuation idea, and the paper will introduce the continuation technique to increase the convergence rate. Therefore, Our continuation log-barrier method for ℓ 1_regularized least square problem is accurate and fast in the sense.

This paper explores a new way to apply knowledge in the area of physics to the artistic technique. Specifically, a new application for imaging techniques in the micrometric scale is proposed: the creation of 3D digital models of small insects which would be robust and precise enough to create physical models up to 10 meters of length while maintaining a realistic appearance. X-ray computed tomography is a technique widely used in the medical world to obtain 3D images of the interior of the body non-intrusively. In this study, micro-computed tomography (microCT) is used to gain a 3D model of a 2.5mm fruit fly. From this image, reverse engineering software is used to create a polygonal model of the sample, which is then repaired until fit for reproduction in a sculpture.

Visual tracking remains a challenging problem because of intrinsic appearance variability of object and extrinsic disturbance. Many algorithms have been recently proposed to capture the varying appearance of targets. Most existing tracking methods, however, fail to estimate the scale and orientation of the target. To deal with this problem, we model the second-order statistics of image regions using a kernel function and perform covariance matching under the Log-Euclidean Riemannian metric. Applying kernel-based covariance matrix as image region descriptor, we construct a region similarity measure that describes the relationship between the candidate object region and a given appearance template. After that, tracking is implemented by minimizing this similarity measure, in which gradient descent method is utilized to iteratively search the best matched object region. The corresponding optimization problem can be derived by calculating the first derivative of the similarity measure with respect to the affine transformation parameters and setting them to be zero. Experimental results compared with several methods demonstrate the robust performance of the proposed algorithm under challenging conditions.

A pulmonary nodules detection algorithm based on 3D template matching method using CT images is proposed. Firstly, lung parenchyma was segmented from CT series images. Secondly, apply the high-pass filter on the lung parenchyma images to enhance the edge of lung nodules. Then, design 3D templates with different size based on the nodules’ features. At last, apply the 3D-SSD (sum of squared differences) template matching algorithm between the 3D templates and the lung fields, and the final matching results were labeled as lung nodules on the original images. Using 20 clinical data set (include 35 pulmonary nodules in 3-20mm) to test the detection method, the accuracy rate is 81.08%, false positive rate (FP) is 5.4% and sensitivity rate is 85.71%. Therefore, the pulmonary nodules detection algorithm proposed in this paper can detect different typological nodules accurately and effectively.

In this work we present an improvement on an iterative binary coloring procedure for image segmentation taken from the literature. We introduce some modifications in the way of dealing with the so-called inconsistent pixels, and we show the results obtained by applying both procedures to a satellite image of the province of Seville. The computational experience that we have performed shows that, in general, the modified procedure leads to images of similar or better quality than the ones obtained by the original procedure, as well as to a significant reduction of the number of final regions.

Moving object detection plays an important role in video surveillance, yet in dynamic scenes it is still a challenging problem. In this work, we develop an efficient algorithm to handling complex dynamic backgrounds by using an adaptive region-based background model. Firstly, the initial background image is partitioned using image over-segmentation methods. Then the input frame is partitioned to image regions according to the obtained partition manner. Features of the image regions are used to construct adaptive mixture Gaussian models. When the background model is updating, the number of component of the mixture Gaussian models is selected adaptively based on the activity level of features. A coarse-to-fine strategy is designed to detect the moving object. The foreground and background are distinguished gradually in region-level and pixel-level through the built background model. Experimental results show that the algorithm proposed in this paper can detect moving objects quickly and effectively.

Imaging model is one of key factors of shape-from-shading (SFS) which reconstruct 3-D shape of objectives from only one image. Most existing SFS methods are based on orthogonal projection. But perspective projection is more accurate than orthogonal projection to simulate the imaging processes of cameras. In this paper, new SFS method under perspective projection with near-scene point lighting source condition is proposed. Imaging model under perspective projection with near-scene lighting source condition is formulated firstly. And the influence of the distance from lighting source to the surface is also considered in the reflectance map equation. Then a set of static Hamilton-Jacobi (H-J) equation is given by the reflectance map equation. The SFS problem is further formulated as viscosity solution of the H-J equation. And the Lax-Friedrichs fast sweeping numerical scheme based on the vanishing viscosity approximation theory is used to solve the H-J equation. At last, experiments on synthetic and real images are performed. Experimental results illustrated the efficiency of the proposed method.

In computer vision as well as in many other domains, high-dimensional data with potentially low intrinsic dimensions need to be modeled. We propose a subspace mixture model under the self-organizing framework, where each neuron learns an adaptive number of dominant eigenvectors. The overall network constitutes a mixture of ordered local subspaces with the advantage of noise smoothing. Experimental results show that the proposed model can accurately represent high-dimensional visual objects such as handwritten digit images subject to substantial variations, and reveal the intrinsic dimensionality of nonlinearly distributed data.

A speech signal being non stationary, the S-transform which combines short-time Fourier analysis and wavelet analysis is an effective tool for analyzing a non stationary signal. Because of the high dimension of time-frequency representations, the singular value decomposition is used to extract the features vectors. In the simulation experiment, firstly, STFT, WT and the S-transform are used to analyze speech signals respectively, and the results show that the time-frequency distribution using the S-transform performs best. Then the right and left singular vectors are extracted using SVD from the time-frequency distributions of Mandarin, Shanghainese, Cantonese and Hokkien as the input feature vectors for SVM to recognize. And the simulation experiment results show that the recognition rate using the S-transform can reach up to 82.5%, higher than using STFT and WT.

The local structure of point cloud is a key problem in point based geometry processing. In this paper, we propose a dictionary learning based method to extract the local structure. The core idea is: As point cloud can be seen as a linear model in local view, we use the union of multi-subspace to approximate it. An overcomplete dictionary D is used to represent the bases of these subspaces. First, we calculate the neighborhood N of each point by k-NN and build EMST on it, marked as T. Then, each edge in T is used to construct a training set. Most of the samples in training set indicate the trend of the point set. At last, we solve the sparse matrix factorization problem recursively to update D until D stops changing. We present 2D/3D experimental results to show that this method can handle manifold/non-manifold structures.

Pyroelectric Infrared (PIR) sensors are excellent devices for wireless sensor network due to its characteristics of low-cost and low-power. PIR sensors are widely used to establish simple but reliable system for detecting targets or triggering alarms. However, processing numerous output data from PIR sensors and correcting the high false generated in the process of classifying and identifying of human targets limit the application scope of PIR sensors. In this paper, a feature extraction and sensor data fusion method to detect and recognize multiple human targets moving in a detection area are presented. Simulation results shows that such approach can reduce computational requirement which indicates that PIR sensors can be used as wireless sensor nodes with limited resources. Additionally, when using the BP neural network, the system can achieve 96% correct identification of individual target data and 90% correct classification of multiple targets mixed data as well.

Voice signals and facial expression changes are synchronized under the different emotions, the recognition algorithm based audio-visual feature fusion is proposed to identify emotional states more accurately. Prosodic features were extracted for speech emotional features, and local Gabor binary patterns were adopted for facial expression features. Two types of features were modeled with SVM respectively to obtain the probabilities of anger, disgust fear, happiness, sadness and surprise, and then fused the probabilities to gain the final decision. Simulation results demonstrate that the average recognition rates of the single modal classifier based on speech signals and based on facial expression reach 60% and 57% respectively, while the multimodal classifier with the feature fusion of speech signals and facial expression achieves 72%.

Object tracking is an important task within the field of computer vision, and the multi-camera tracking with disjoint view is more applicable. This paper focus on introducing human memory mechanism into multi-camera human tracking problem that the Field Of View (FOV) of cameras are not necessarily overlapping, and proposing a camera handover scheme based on memory. In the modeling process, every target goes through transmission and storage of three spaces: sensory memory, short-term memory and long-term memory. After learning, memory-based target handoff can remember target appeared earlier, and when faced with similar goals, it can extract and activate the target in memory in time, so that it can quickly achieve object handoff between different camera tracking. Preliminary experiments show that this scheme is effect in camera handover and multi-camera human tracking.

A face recognition algorithm is proposed based on Curvelet transform and Class estimated basis Space singular Value De-composition (CSVD). Face images are decomposed by using Curvelet transform firstly. As a result, Curvelet coefficients in different scales and various angels are obtained. Then, the images reconstructed by the Curvelet coefficients of the coarse layer are processed by a Fourier transform with invariant prosperity against spatial translation. CSVD algorithm is used to reduce the dimensionality and extract the feature of the amplitude spectrum face. Finally, the nearest neighbor decision rule is applied to identify the unknown face. The standard face databases of ORL, FERET and Yale are selected to evaluate the recognition accuracy of the algorithm. The results show that the proposed algorithm is used to improve the recognition rate effectively.

A video-based fire detection system for outdoor surveillance should be capable of continuous operation under various environmental conditions, such as the change of background illumination and variation of background objects. In this paper, we present a novel background modeling method in which a nonparametric statistical test is utilized for effective and robust detection of stationary background corners. The practical value of the method is demonstrated with an oil-field surveillance system where it is applied for video-based fire detection. The validation results and analysis indicate that the proposed method is able to cope with small occlusions and periodic motions.

To explore a concise and efficient approach to address video semantic retrieval problem, we creatively propose a concept of structured semantic string to realize video indexing and retrieval in this paper. Integrated with the technology of NLP (Natural Language Processing) and documents inverted index, our method could help represent information on sentence level so that computer could better understand videos’ descriptions. In addition, expansion on synonyms and hypernyms contributes to comprehensive semantic expressing. Presently, the approach has been applied in our distributed video semantic retrieval system – Xunet, and it is proved to achieve high recall rate and precision rate.

This paper proposes a color image segmentation algorithm based on affinity propagation clustering. Firstly, the color image is converted to other color space; and then color image is sampled, and the sampling data is clustered by affinity propagation algorithm with given the number, then we can get their class label of the rest data according to the maximum similarity rules; finally, using the regional combined methods ,we can remove independent small regions, and get revised image segmentation result. The simulation experiment shows that this method can achieve the application of affinity propagation algorithm in large-scale color image segmentation. And it has a faster processing speed and satisfactory segmentation results, more in line with the human whole visual perception.

Monitoring the changes of vigilance is very important, because decline in the operator’s alert is one of the primary causal factors for many accidents. Electroencephalogram (EEG) based brain computer interface (BCI) for vigilance analysis and estimate have been rapidly advanced in recent years. However, there still exist many difficulties and challenges in their application. In this paper, we propose an applicable EEG-based BCI system for vigilance monitoring: the number of record channels could be reduced greatly with considering individual variability in EEG, since the practical BCI is expected to be implemented with a small number of channels. Furthermore, the vigilance model can be constructed according to various application situation. With the dry electrodes we are now developing, the system will be conveniently applied to vigilance monitoring in the future.