Joint factor analysis (JFA) is the mainstream algorithm in the text-independent speaker verification systems due to its clear method of modeling the spaces. However, the inevitable overlaps between the speaker space and the channel space obtained by JFA are caused because of the limitations of the algorithm process. To resolve this problem, the space transformation based on the signal subspace is proposed. Compared with JFA algorithm without the space transformation, an equal error rate (EER) reduction of 9.2% is obtained on the telephone section of the core condition trials of the NIST SRE 2008.

To improve the convergence performance of differential evolution algorithm, an adaptive multiple strategy differential evolution algorithm (AMSDE) with guiding scheme of Pbest is proposed. The library of control parameters of the crossover probability, the library of scale parameters of the mutation and the library of the differential mutation strategy are designed in AMSDE. Thereinto, the crossover probability is generated by Logistic chaotic systems, the scale parameter is produced by means of a linear changing scheme, and the library of differential mutation strategy consists of six widely used differential mutation strategies. Subsequently, the framework of AMSDE is given. Finally, simulation results on 25 benchmark test functions demonstrate that AMSDE achieves better convergence precision and a higher convergence speed. And AMSDE outperforms the two state-of-the-art variants of differential evolution algorithms, JADE and CoDE.

Human behavior recognition is a hot issue in computer vision. However, most of the existing algorithms only use RGB or depth video sequence, and few of them are combined for behavior recognition. Due to their own advantages and complementary information, the characteristics of depth images and RGB images are studied, and two kinds of robust descriptors and some fusion schemes for them are proposed in this paper. Then, the support vector machine classifiers with different kernels are adopted. Results of extensive experiments on the challenging DHA dataset show that the accuracies of the proposed descriptors are higher than those of the state-of-the-art algorithms. Meanwhile, the performance of the algorithm with the combination of depth information and RGB is improved, and it is better than that of the algorithm with sole descriptor. Moreover, the proposed descriptors have strong robustness, discriminability and stability.

Current inference algorithms of Bayesian networks are weak on inference precision and inference time to a certain degree. Therefore, in this paper a practical and reliable inference method, samplingapproximate inference algorithm based on Markov blanket (LSIA-MB), is presented. Firstly, HITON_MB algorithm is utilized to obtain the Markov blanket of the query node and then the dynamic programming algorithm is used to learn the posterior probability of edges to get a Markov local network model of the query node.Finally, Gibbs sampling inference algorithm is executed on the Markov local model. The sampling on the local model significantly reduces the calculation dimensions. The inference precision is retained because the Markov local model contains the complete information associated with the query node. Algorithm analysis and experimental results on standard Alarm network show LSIA-MB algorithm significantly reduces the inference time and improves the inference precision. The inference results of LSIA-MB algorithm on the Shanghai stock exchange network show the algorithm has strong practicability.

All of the finite interpretations of first-order languages are considered in layers according to their cardinalities for the first time, and n-truth degrees of formulae under each layer of the class of interpretations are proposed.Then, the definition of the vector truth degrees of formulae is given to describe the truth degrees of first-order formulae more intuitively and more accurately. Moreover, some basic properties are proved, and it is pointed out that the vector truth degrees of formulae preserve Modus Ponens(MP) rule, Hypothetical Syllogism(HS) rule and rule of generalization(Gen) of the formal reasoning in the predicate logic. Thus, a kind of possible frame for approximate reasoning in predicate logic system is provided.

Truth table reduction simplifies the analysis and the design of the digital logic circuit. It is also used for the judgment of the propositional logic value and the equivalence relation of the compound statement in the artificial intelligence theory. A granular matrix based rapid reduction algorithm for multivariable truth table is proposed in this paper by defining the truth table as the logic information system, and the granular matrix is introduced to describe the knowledge in different granular spaces. Then, the truth table reduction is converted to the attribute and the attribute value reduction of the logic information system implemented by granular matrix computation. As an example, the proposed algorithm is applied to the design of the light emitting diode seven segment display. The experimental results show high efficiency of the proposed method.

Image segmentation based on active contour model(ACM)is a hotspot in computer vision. Firstly, the mathematical model of ACM and the related numerical implementation are discussed. Secondly, two development models of ACM are summarized in contrast: parameter active contour model and geometry contour model. The comparative study of these two kinds of models and the deep discussion on the scope of image segmentation application and advantages and disadvantages of each kind of techniques are carried out. Finally, the directions of future development in ACM are pointed out.

The difference of feature data range is ignored when Euclidean distance is used as a vector similarity metric. And thus, the classification accuracies of the traditional learning vector quantization algorithm (LVQ) and its variants are affected. To solve the problem, a vector similarity metric is proposed and based on this metric and generalized LVQ(GLVQ), an algorithm, GLVQ-Range, is put forward. The classification accuracy and the computation speed of the proposed algorithm are tested on 8 datasets of UCI machine learning repository, compared with those of the traditional alternative LVQ algorithms. The practicability of the proposed algorithm in real production environment is verified on the video vehicle classification dataset.

The local binary pattern(LBP) descriptors employed in the content-based image retrieval system lack the abilities to describe the spatial relationships and have longer dimension of feature vector. In this paper, an improved LBP (ILBP) texture descriptor based on spatial statistical feature of LBP code pair is proposed. The original image is converted to the LBP pseudo image using LBP coding method for micro-pattern, and then several statistics of LBP code pair are extracted to form the feature vector for describing the texture attributes of images. Experiments are preformed on the content-based image retrieval prototype platform. Experimental results show that compared with other LBP descriptors the ILBP descriptor further enhances the description ability of LBP descriptor and substantially reduces the feature vector dimension with better query accuracy and query efficiency.

Representation, relation and operation of granules are the main research content of granular computing. A hyperbox granule is represented by a vector including a beginning point and an end point. The inconsistency between the partial ordering relation in vector space and the partial ordering relation in hyperbox granule space is analyzed and then eliminated by the order-preserving function. The fuzzy inclusion relation between two hyperbox granules is formed by nonlinear positive valuation function and the order-preserving function between the lattice and its dual lattice. The join operator and decomposition operator between two granules are designed to achieve the granules with different granularity. The algebraic system which is composed of hyperbox granule set, fuzzy inclusion relation and the operators between two granules is proved as fuzzy lattice. Hyperbox granular computing classifiers are formed based on fuzzy lattice, and verified by classification problems on machine learning dataset. The experimental results show that hyperbox granular computing classifiers have a generalization ability comparable to that of fuzzy lattice reasoning classifiers with less number of hyperbox granules.

Utilizing the characteristic of cloud model principles which can make good balance between the randomness and the fuzziness, an adaptive quantum-behaved particle swarm optimization algorithm based on cloud model is proposed. Firstly, the control mechanism of quantum-behaved particle swarm optimization algorithm is analyzed. On this basis, the absorption-expansion factor of each particle is adaptively controlled by cloud operators to achieve the dynamic adjustment to the positions of particles in evolutionary process. Thus, the proposed algorithm obtains a higher convergence speed and a stronger global search ability. Programs are modified for the targeted optimization to make the proposed algorithm effectively avoid falling into local optimum. The results of simulation experiments with typical test functions show that the proposed algorithm has advantages in search ability, accuracy and stability, and it is more effective than other similar algorithms.

To gain a fast, accurate and parsimonious fuzzy neural network, an effective incremental sequential learning algorithm for parsimonious fuzzy neural networks (ISL-FNN)is proposed. The pruning strategy is introduced into the generation of neurons. The error reduction ratio is used to define the influence of input data on the output and the influence is utilized for the generation of neurons. In the parameter learning phase, all the free parameters of hidden units, including the newly created and the originally existing, are updated by the extended Kalman filter method. The performance of ISL-FNN is compared with several existing algorithms on some benchmark problems. Result indicates that ISL-FNN produces similar or even better accuracies with less number of rules.