Artificial Intelligence and Soft Computing

A parallel architecture of the Levenberg-Marquardt algorithm for training a feedforward neural network is presented. The proposed solution is based on completely new parallel structures to effectively reduce high computational load of this algorithm. Detailed parallel neural network structures are explicitely discussed.

The paper presents the application of our clustering technique based on generalized self-organizing neural networks with evolving tree-like structures to complex cluster-analysis problems including, in particular, the sample-based and gene-based clusterings of microarray

Leukemia

gene data set. Our approach works in a fully unsupervised way, i.e., without the necessity to predefine the number of clusters and using unlabelled data. It is particularly important in the gene-based clustering of microarray data for which the number of gene clusters is unknown in advance. In the sample-based clustering of the

Leukemia

data set, our approach gives better results than those reported in the literature and obtained using a method that requires the cluster number to be defined in advance. In the gene-based clustering of the considered data, our approach generates clusters that are easily divisible into subclusters related to particular sample classes. It corresponds, in a way, to subspace clustering that is highly desirable in microarray data analysis.

This paper presents a new concept of representation of data and their relations in neural networks which allows to automatically associate, reproduce them, and generalize about them. It demonstrates an innovative way of developing emergent neural representation of knowledge using a new kind of neural networks whose structure is automatically constructed and parameters are automatically computed on the basis of plastic mechanisms implemented in a new associative model of neurons - called as-neurons. Inspired by the plastic mechanisms commonly occurring in a human brain, this model allows to quickly create associations and establish weighted connections between neural representations of data, their classes, and sequences. As-neurons are able to automatically interconnect representing similar or sequential data. This contribution describes generalized formulas for quick analytical computation of the structure and parameters of ANAKG neural graphs for representing and recalling of training sequences of objects.

Complexity of shallow (one-hidden-layer) networks representing finite multivariate mappings is investigated. Lower bounds are derived on growth of numbers of network units and sizes of output weights in terms of variational norms of mappings to be represented. Probability distributions of mappings whose computations require large networks are described. It is shown that due to geometrical properties of high-dimensional Euclidean spaces, representation of almost any randomly chosen function on a sufficiently large domain by a shallow network with perceptrons requires untractably large network. Concrete examples of such functions are constructed using Hadamard matrices.

In this paper, we present new probabilistic neural network (PNN) training procedure for classification problems. Proposed procedure utilizes the State-Action-Reward-State-Action algorithm (SARSA in short), which is the implementation of the reinforcement learning method. This algorithm is applied to the adaptive selection and computation of the smoothing parameter of the PNN model. PNNs with different forms of the smoothing parameter are regarded. The prediction ability for all the models is assessed by computing the test error with the use of a 10-fold cross validation (CV) procedure. The obtained results are compared with state-of-the-art methods for PNN training.

Paper considers three Hopfield based architectures in the stereo matching problem solving. Together with classical analogue Hopfield structure two novel architectures are examined: Hybrid-Maximum Neural Network and Self Correcting Neural Network.Energy functions that are crucial for the network performance and working algorithm are also presented.All considered structures are tested to compare their performance features. Two of them are particularly important: accuracy and computational time. For the experiment real and simulated stereo images are used. Obtained results lead to the conclusion about feasibility of considered architectures in the stereo matching problem solving for real time applications.

The present article puts forward a completely new technology development , a spin glass-like molecular implementation of the Hopfield neural structure. This novel approach uses magnetic molecules homogenously distributed in mesoporous silica matrix, which forms a base for a converting unit, an equivalent of a neuron in the Hopfield network. Converting units interact with each other via a fully controlled magnetic fields, which corresponds to weighted interconnections in the Hopfield network. This novel technology enables building fast, high-density content addressable associative memories. In particular, it is envisaged that in the future this approach can be scaled up to mimic memory with human-like characteristics. This would be a breakthrough in artificial brain implementations and usher in a new type of highly intelligent beings. Another application relates to systems designed for multi-objective optimization (multiple criteria decision making).

This paper presents the application of some Computational Intelligence methods for obtaining a classifier analysing employees to form work groups. The proposed bio-inspired solution analyses employees using data gathered from their professional attitudes and skills, then suggests how to form groups of human resources within a company that can effectively work together. The same proposed tool provides employers with a fair and effective means for employee evaluation. In our approach, employee profiles are processed by a dedicated Radial Basis Probabilistic Neural Network based classifier, which finds non-explicit custom-created groups. The accuracy of the classifier is very high, revealing the potential efficacy of the proposed bio-inspired classification system.

In the paper, the parallel realization of the Boltzmann Restricted Machine (RBM) is proposed. The implementation intends to use multicore architectures of modern CPUs and Intel Xeon Phi coprocessor. The learning procedure is based on the matrix description of RBM, where the learning samples are grouped into packages, and represented as matrices. The influence of the package size on convergence of learning, as well as on performance of computation, are studied for various number of threads, using conventional CPU and Intel Phi architecures. Our research confirms a potential usefulness of MIC parallel architecture for implementation of RBM and similar algorithms.

Automatic software parallelization is a key issue for high performance computing. There are many algorithms to transform program loop nests to multithreaded code. However, the time of a transformation process is usually unknown, especially for transitive closure based algorithms. The computational complexity of transitive closure calculation algorithms is relatively high and may prevent applying corresponding transformations. The paper presents the prediction of loop transformation time by means of an artificial neural network for the source-to-source TRACO compiler. The analysis of a loop nest structure and dependences is used to estimate the time of TRACO transformations. The training of a Feed-Forward Neural Network is used to make a decision about transformation time. Experiments with various NAS Parallel Benchmarks show promise for the use of neural networks in automatic code parallelization and optimization.

This paper presents a new concept of a dual neural network which is hybrid of linear and nonlinear network. This approach allows for solving the problem of Parity-3 with only one sigmoid neuron or Parity-7 with 2 sigmoid neurons that is shown in the analytical and experimental manner. The paper describes the architecture of ANN, presents an analytical way of choosing the weights and the number of neurons, and provides the results of network training for different ANN architectures solving the Parity-

N

problem.

Regularization networks represent a kernel-based model of neural networks with solid theoretical background and a variety of learning possibilities. In this paper, we focus on its extension with multi-kernel units. In particular, we describe the architecture of a product unit network, and we propose an evolutionary learning algorithm for setting its parameters. The algorithm is capable to select different kernels from a dictionary and to set their parameters, including optimal split of inputs into individual products. The approach is tested on real-world data from sensor networks area.

Our investigation applies a fuzzy grouping model in order to identify potential enterprise clusters based on their characteristic manufacturing activities in a specific city. The aim is to create clusters towards the construction of competitive advantages, cost reduction and economies of scale. We utilize tools of Fuzzy Sets Theory, evaluating productive capacities of local enterprises under Moore Families. Results conclude in 16 different clusters formed by 2, 3, 4 and 5 firms located in 6 different zones of a specific city. This work seeks to shed light in the conformation of groups under uncertain conditions, and the deep examination of the manufacturing activities in a specific territory for decision and policy making.

This paper proposes a global fingerprint feature named QFingerMap that provides fuzzy information about a fingerprint image. A fuzzy rule that combines information from several QFingerMaps is employed to register an individual in a database. Error and penetration rates of a fuzzy retrieval system based on those rules are similar to other systems reported in the literature that are also based on global features. However, the proposed system can be implemented in hardware platforms of very much lower computational resources, offering even lower processing time.

Fuzzy sets and rough sets are well-known approaches to incomplete or imprecise data. In the paper we compare two formalizations of these sets within one of the largest repositories of computer-checked mathematical knowledge – the Mizar Mathematical Library. Although the motivation was quite similar in both developments, these approaches – proposed by us – vary significantly. Paradoxically, it appeared that fuzzy sets are much closer to the set theory implemented within the Mizar library, while in order to make more feasible view for rough sets we had to choose relational structures as a basic framework. The formal development, although counting approximately 15 thousand lines of source code, is by no means closed – it allows both for further generalizations, building on top of the existing knowledge, and even merging of these approaches. The paper is illustrated with selected examples of definitions, theorems, and proofs taken from rough and fuzzy set theory formulated in the Mizar language.

This paper is aimed at the giving of a comparative approach to the preferences modelling. This approach is conceived to grasp the fuzzy nature of preferences what determines the choice two fuzzy logic formalisms for their representation discussed by P. Hajek and L. Godo. These two (appropriately modified) formalism are used to propose two formalism for preferences:

Fuzzy Modal Preferential Logic

(FMPL) and

Comparative Possibilistic Multi-Modal Propositional Logic

(CPMPL). We also justify some metalogical properties of both systems such as their completeness and we discuss a satisfiability problem for them. In result, we propose a short juxtaposition of the properties of the considered systems.

Imperfect methods of aquiring knowledge from experts in order to create fuzzy rules are generally known [16,4,25]. Since this is a very important part of fuzzy inference systems, this article focuses on presenting new learning methods for fuzzy rules. Referring to earlier work, the authors extended learning methods for fuzzy rules on applications of Type-2 fuzzy logic systems to control filters reducing air pollution. The filters use Selective Catalytic Reduction (SCR) method and, as for now, this process is controlled manually by a human expert.

In this paper, some results concerning analytical methods of fuzzy modeling, especially so called P1-TS fuzzy rule-based systems are described. The basic notions and facts concerning the theory of fuzzy systems are briefly recalled, including a method for overcoming or at least weakening the curse of dimensionality. A P1-TS system performing the function of the fuzzy JK flip-flop, as well as optimal controller for the 2nd order dynamical plant are described. Next, we show how to use the idea of P1-TS system for identification of some class of nonlinear dynamical systems. We briefly characterize FPGA hardware implementation of the P1-TS system. A result of a mobile robot navigation system design is described, as well. Finally, we show how to obtain a highly interpretable fuzzy classifier as a medical decision support system, by using both the theory of P1-TS system with a large number of inputs in conjunction with the idea of meta-rules, and gene expression programming method.

In this paper, we describe

fuzzy agglomerative clustering

, a brand new fuzzy clustering algorithm. The basic idea of the proposed algorithm is based on the well-known hierarchical clustering methods. To achieve the soft or fuzzy output of the hierarchical clustering, we combine the single-linkage and complete-linkage strategy together with a

fuzzy distance

. As the algorithm was created recently, we cover only some basic experiments on synthetic data to show some properties of the algorithm. The reference implementation is freely available.

Entropy of an intuitionistic fuzzy set (IFS) is used to indicate the degree of fuzziness for IFSs. In this paper we deal with entropies of IFSs. We first review some existing entropies of IFSs and then propose a new entropy measure based on exponential operations for an IFS. Finally, comparisons are made with some existing entropies to show the effectiveness of our proposed one.

The paper presents the Characteristic Objects method as a potential multi-criteria decision-making method for use in medical issues. The proposed approach is compared with TOPSIS and AHP. For this purpose, assessment of the severity of Chronic Liver Disease (CLD) is used. The simulation experiment is presented on the basis of the Model For End-Stage Liver Disease (MELD). The United Network for Organ Sharing (UNOS) and Eurotransplant use MELD for prioritizing allocation of liver transplants. MELD is calculated from creatinine, bilirubin and international normalized ratio of the prothrombin time (INR). The correctness of the selection is examined among randomly selected one million pairs of patients. The result is expressed as a percentage of agreement between the assessed method and MELD selection. The Characteristic Objects method is completely free of the rank reversal phenomenon, obtained by using the set of characteristic objects. In this approach, the assessment of each alternative is obtained on the basis of the distance from characteristic objects and their values. As a result, correctness of the selection obtained by using the Characteristic Objects method is higher than those obtained by TOPSIS or AHP techniques.

The paper presents a simple method of CwW that is based on RDM-models of intervals and on the multidimensional RDM-interval arithmetic. The method is explained on the example of popular Zadeh’s challenge problem known as ‘Balls in a box’ problem. On the base of presented calculations, the general methodology of solving CwW problems with the application of simplified RDM-models of quantifiers is formulated.

The Ordered Fuzzy Numbers (OFN) were defined over 10 years ago as a tool for processing fuzzy numbers. This model has an additional feature used in processing, namely direction. It allows to define arithmetical operations in a new way. Proposed methods retain the basic computational properties of the operations known for the real numbers. Apart from a good calculations, OFNs also offer new possibilities for processing imprecise information. The new property - a direction - has a major impact on the calculations, but gives also a new potential for processing data in the fuzzy systems. We can include to the fuzzy value some more interpretation than membership value. If we want take into account this additional information in the processing of fuzzy system, we need the methods which are sensitive for the direction.

This publication presents the basic tool for processing a fuzzy statement ’

A

is

B

’ where

A

and

B

are OFNs. It can be called a compatibility of

A

with

B

. New proposition is sensitive for the direction and bases on the conception of the

Direction Determinant

proposed in previous studies on the topic.

This paper focuses on problems related to learning rules using numerical data for the

Hierarchical Fuzzy Logic Systems (HFLS)

described in [9]. Learning rules for

Fuzzy Logic Systems (FLS)

or

Fuzzy Controller (FC)

in short could be accomplished by using many different approaches, building one, complex rulebase using all available input and output variables. Using hierarchical structure we could avoid this problem by problem division into subproblems with smaller dimensions. ”Hierarchical” means that fuzzy sets produced as output of one of fuzzy controllers are then processed as an inputs of another one as the sets of auxiliary variables. The main problem is to learn rulebase with numerical data, which does not contain any data for those auxiliary variables. The main scope of this paper is to present an algorithm that is being a solution to this problem and provides support for selective activation of unit FC. The proposal presented in this paper operates on a type-1 HFLS, built with the fuzzy controllers (in the sense of Mamdani). An example of single-player games, i.e. where the “enemy” is controlled by agents is used.

In this paper, a new approach to the rule-base evidential reasoning (

RBER

) based on a new formulation of fuzzy rules is presented. We have shown that the traditional fuzzy logic rules lose an important information when dealing with the intersecting fuzzy classes, e.g., such as

Low

and

Medium

, and this property may lead to the controversial results. In the framework of our approach, an information of the values of all membership functions representing the intersecting (competing) fuzzy classes is preserved and used in the fuzzy logic rules. As

RBER

methods are based on the synthesis of fuzzy logic and the Dempster-Shafer theory of evidence (

DST

), the problem of the combination of basic probability assignments (

bpa

s) from different sources of evidence arises. The classical Dempster’s rule of combination is usually used for this purpose. The classical Dempster’s rule may provide controversial results in the case of great conflict and is not idempotent one. We show that the Dempster’s rule may provide unreasonable results not only in the case of large conflict, but in the case of complete absence of conflict, too. At the end, we show that in the cases of small and large conflict, the use of simple averaging rule for combination of

bpa

s seems to be a best choice. The developed approach is illustrated by the solution of simple, but real-world problem of diagnostics of type 2 diabetes.

In fuzzy clustering, the fuzzy c-means (FCM) algorithm is the most commonly used clustering method. However, the FCM algorithm is usually affected by initializations. Incorporating FCM into switching regressions, called the fuzzy c-regressions (FCR), has also the same drawback as FCM, where bad initializations may cause difficulties in obtaining appropriate clustering and regression results. In this paper, we proposed the bias-correction fuzzy c-regressions (BFCR) algorithm by incorporating bias-correction FCM (BFCM) into switching regressions. Some numerical examples were used to compare the proposed algorithm with some existing fuzzy c-regressions methods. The results indicated the superiority and effectiveness of the proposed BFCR algorithm.

In this paper a new interval Type-2 fuzzy neural network will be presented for function approximation. The proposed neural network is based on Locally Linear Model Tree (LOLIMOT) which is a fast learning algorithm for Locally Linear Neuro-Fuzzy Models (LLNFM). In this research, main measures are to be robust in presence of outlier data and be fast in refining steps. The proposed combination between LOLIMOT learning algorithm and interval type 2 fuzzy logic systems presents a good performance both in robustness and speed measures. The results show that the proposed method has good robustness in presence of noise as we can see in experiments conducted using corrupted data. Also this method has eligible speed as it can be seen in the results.

The paper studies the problem of continuous function optimization. Proposed are twelve hybrids of known methods such as Particle Swarm Optimization (also in a two-subpopulation version), quasi-Newton method and Nelder-Mead method. Described modifications are introduced in order to improve performance and increase the accuracy of known methods. Algorithms are tested against eight benchmark functions and compared with classical versions of: Particle Swarm Optimization algorithm, Newton’s, quasi-Newton and Nelder-Mead methods. Presented results allow to indicate two methods that perform satisfactorily in most cases.

The work considers the problem of three-dimensional bin packaging (3D-BPP), where a load of maximum volume is put in a single container. To solve the above mentioned problem there was a coevolutionary parallel algorithm used basing on the separate evolution of cooperating subpopulations of possible solutions. Computational experiments were conducted in a neighbourhood of clusters and aimed to examine the impact of parallelization algorithm on the computation time and the quality of the obtained solutions.

Possible improvement of a successful adaptive SHADE variant of differential evolution is addressed. Exploitation of exponential crossover was applied in two newly proposed SHADE variants. The algorithms were compared experimentally on CEC 2013 test suite used as a benchmark. The results show that the variant using adaptive strategy of the competition of two types of crossover is significantly more efficient than other SHADE variants in 7 out of 28 problems and not worse in the others. Thus, this SHADE with competing crossovers can be considered superior to original SHADE algorithm.

One of the important and still not fully addressed issues in evolving decision trees is the induction time, especially for large datasets. In this paper, the authors propose a parallel implementation for Global Decision Tree system that combines shared memory (OpenMP) and message passing (MPI) paradigms to improve the speed of evolutionary induction of decision tree. The proposed solution is based on the classical master-slave model. The population is evenly distributed to available nodes and cores, and the time consuming operations like fitness evaluation and genetic operators are executed in parallel on slaves. Only the selection is performed on the master node. Efficiency and scalability of the proposed implementation is validated experimentally on artificial datasets. It shows noticeable speedup and possibility to efficiently process large datasets.

This paper discusses selected aspects of evolutionary search algorithms guided by grammars, such as Grammar Guided Genetic Programming or Grammatical Evolution. The aim of the paper is to demonstrate that, when the efficiency of the search process in such environment is considered, it is not only the language defined by a grammar that is important, but also the form of the grammar plays a key role. In the most common current approach, the person who sets up the search environment provides the grammar as well. However, as demonstrated in the paper, this may lead to a sub-optimal efficiency of the search process. Because an infinite number of grammars of different forms can exist for a given language, manual construction of the grammar which makes the search process most effective is generally not possible. It seems that a desirable solution would be to have the optimal grammar generated automatically for the provided constrains. This paper presents possible solutions allowing for automatic grammar induction, which makes the search process more effective.

In this paper we study the search abilities of binary one-point crossover (1ptc) operator in a genetic algorithm (GA). We show, that under certain conditions, GA is capable of using only a 1ptc operator to explore the entire search space, fighting premature convergence. Further, we prove that to restore the entire space from any two binary chromosomes, each of length

n

, at least 2

n

 − 1

 − 1 one-point crossover operations is needed. This number can serve as a measure for comparing the search speed of the different algorithms. Moreover, we propose an algorithm spanning the search space in the minimal number of crossovers.

In this study the promising Multiple-choice strategy for PSO (MC-PSO) is enhanced with the blind search based single dimensional mutation. The MC-PSO utilizes principles of heterogeneous swarms with random behavior selection. The performance previously tested on both large-scale and fast optimization is significantly improved by this approach. The newly proposed algorithm is more robust and resilient to premature convergence than both original PSO and MC-PSO. The performance is tested on four typical benchmark functions with variety of dimension settings.

In this paper a new three level, hybrid optimization method is proposed. Differential evolution is hybridized with traditonal gradient optimization. Some ideas from simulated annealing are also employed. Usefulness of the proposed method is supported by numerical simulations.

This research deals with the hybridization of the two softcomputing fields, which are chaos theory and evolutionary computation. This paper aims on the experimental investigations on the chaos-driven evolutionary algorithm Differential Evolution (DE) concept. This research represents the continuation of the satisfactory results obtained by means of chaos embedded (driven) DE, which utilizes the chaotic dynamics in the place of pseudorandom number generators This work is aimed at the tuning of the complex chaotic dynamics directly injected into the DE. To be more precise, this research investigates the influence of different parameter settings for discrete chaotic systems to the performance of DE. Repeated simulations were performed on the IEEE CEC 13 benchmark functions set in dimension of 30. Finally, the obtained results are compared with canonical DE and jDE.

In this paper we investigate possibilities of relevance learning in learning matrix quantization and discuss their mathematical properties. Learning matrix quantization can be seen as an extension of the learning vector quantization method, which is one of the most popular and intuitive prototype based vector quantization algorithms for classification learning. Whereas in the vector quantization approach vector data are processed, learning matrix quantization deals with matrix data as they occur in image processing of gray-scale images or in time-resolved spectral analysis. Here, we concentrate on the consideration of relevance learning when learning matrix quantization is based on the Schatten-

p

-norm as the data dissimilarity measure. For those matrix systems exist more relevance learning variants than for vector classification systems. We contemplate several approaches based on different matrix products as well as tensor operators. In particular, we discuss their mathematical properties related to the relevance learning task keeping in mind the stochastic gradient learning scheme for both, prototype as well as relevance learning.

Active Learning (AL) is an emerging field of machine learning focusing on creating a closed loop of learner (statistical model) and oracle (expert able to label examples) in order to exploit the vast amounts of accessible unlabeled datasets in the most effective way from the classification point of view.

This paper analyzes the problem of multiclass active learning methods and proposes to approach it in a new way through substitution of the original concept of predefined utility function with an ensemble of learners. As opposed to known ensemble methods in AL, where learners vote for a particular example, we use them as a black box mechanisms for which we try to model the current competence value using adaptive training scheme.

We show that modeling this problem as a multi-armed bandit problem and applying even very basic strategies bring significant improvement to the AL process.

The article concerns of the problem of regression functions estimation when the output is contaminated by additive nonstationary noise. We investigate the model

$y_i = R\left( {{\bf x _i}} \right) + Z _i ,\,i = 1,2, \ldots n$

, where

x

i

is assumed to be the set of deterministic inputs (

d

-dimensional vector),

y

i

is the scalar, probabilistic outputs, and

Z

i

is a measurement noise with zero mean and variance depending on

n

.

$R\left( . \right)$

is a completely unknown function. The problem of finding function

$R\left( . \right)$

may be solved by applying non-parametric methodology, for instance: algorithms based on the Parzen kernel or algorithms derived from orthogonal series. In this work we present the orthogonal series approach. The analysis has been made for some class of nonstationarity. We present the conditions of convergence of the estimation algorithm for the variance of noise growing up when number of observations is tending to infinity. The results of numerical simulations are presented.

An application of Ground Penetrating Radar to landmine detection is presented. Using our prototype GPR system, we collect high-resolution 3D images, so called C-scans. By sampling 3D windows from C-scans, we generate large data sets for learning. We focus on experimentations with different recipes for growing shallow decision trees under the real-boost procedure. A particular attention is paid to the exponential criterion working as impurity function, in comparison to well known impurities. In the light of a theoretical bound on true error, driven from the properties of boosting, we check how greedy learning approaches translate in practice (for our GPR data) onto test error measures.

In this paper a new approach for construction of neuro-fuzzy systems for nonlinear classification is introduced. In particular, we concentrate on the flexible neuro-fuzzy systems which allow us to extend notation of rules with weights of fuzzy sets. The proposed approach uses possibilities of hybrid evolutionary algorithm and interpretability criteria of expert knowledge. These criteria include not only complexity of the system, but also semantics of the rules. The approach presented in our paper was tested on typical nonlinear classification simulation problems.

The basic nearest neighbour algorithm has been designed to work with complete data vectors. Moreover, it is assumed that each reference sample as well as classified sample belong to one and the only one class. In the paper this restriction has been dismissed. Through incorporation of certain elements of rough set and fuzzy set theories into

k

-nn classifier we obtain a sample based classifier with new features. In processing incomplete data, the proposed classifier gives answer in the form of rough set, i.e. indicated lower or upper approximation of one or more classes. The basic nearest neighbour algorithm has been designed to work with complete data vectors and assumed that each reference sample as well as classified sample belongs to one and the only one class. Indication of more than one class is a result of incomplete data processing as well as final reduction operation.

Cross-entropy clustering (CEC) is a density model based clustering algorithm. In this paper we apply CEC to the one-class classification, which has several advantages over classical approaches based on Expectation Maximization (EM) and Support Vector Machines (SVM). More precisely, our model allows the use of various types of gaussian models with low computational complexity. We test the designed method on real data coming from the monitoring systems of wind turbines.

Extraction and detailed analysis of sound files using the MPEG 7 standard descriptors is extensively explored. However, an automatic description of the specific field of sounds of nature still needs an intensive research. This publication presents a comparison of effectiveness of time and frequency descriptors applied in recognition of species of birds by their voices. The results presented here are a continuation of the research/studies on this subject. Three different conceptions of classification - the WEKA system as classical tool, linguistically modelled fuzzy system and artificial neural network were used for testing the descriptors’ effectiveness. The analysed sounds of birds come from 10 different species of birds: Corn Crake, Hawk, Blackbird, Cuckoo, Lesser Whitethroat, Chiffchaff, Eurasian Pygmy Owl, Meadow Pipit, House Sparrow and Firecrest. For the analysis of the physical features of a song, MPEG 7 standard audio descriptors were used.

In this paper, a mechanical imbalance prediction problem for a milling tool heads used in Computer Numerical Control (CNC) machines was studied. Four classes of the head imbalance were examined. The data set included 27334 records with 14 features in the time and frequency domains. The feature selection procedure was applied in order to extract the most significant attributes. Only 3 out of 14 attributes were selected and utilized for the representation of each signal. Seven computational intelligence methods were applied in the prediction task: K–Means clustering algorithm, probabilistic neural network, single decision tree, boosted decision trees, multilayer perceptron, radial basis function neural network and support vector machine. The accuracy, sensitivity and specificity were computed in order to asses the performance of the algorithms.

Rice is an important crop utilized as a staple food in many parts of the world and particularly of importance in Asia. The process to grow rice is very human labor intensive. Much of the difficult labor of rice production can be automated with intelligent and robotic platforms. We propose an intelligent agent which can use sensors to automate the process of distinguishing between rice and weeds, so that a robot can cultivate fields. This paper describes a feature-based learning approach to automatically identify and distinguish weeds from rice plants. A Harris Corner Detection algorithm is firstly applied to find the points of interests such as the tips of leaf and the rice ear, secondly, multiple features for each points surrounding area are extracted to feed into a machine learning algorithm to discriminate weed from rice, last but not least, a clustering algorithm is used for noise removal based on the points position and density. Evaluation performed on images downloaded from internet yielded very promising classification result.

This article focuses on calculation of how close the estimated average error of Prolate Spheroidal Wave Functions (PSWFs) approaches the actual radius of information in the presence of both jitter and approximation and quantization error. The existing upper bound of the estimation is here paired with lower bound and compared to bounds on information radius. Performed calculation are to obtain the precision of error estimation.

The paper presents a formal approach to developing new heuristic methods for finding solutions of discrete optimization problems. The presented approach is based on algebraic-logical meta-model of multistage decision process (ALMM of DMP) that has been developed by the author. Definitions are provided for two deterministic classes of multistage decision processes: common multistage decision processes (cMDP) and multistage dynamic decision processes (MDDP). The paper presents some part of research results pertaining to heuristic methods utilising ALMM of MDP. It lays out a three stage concept of heuristic method synthesis involving local optimization together with two heuristic methods based on the said concept: Machine Learning Based on ALMM of DMP and the Substitute Task Method.

Here, a comparative study of information fusion methods for instance object detection is proposed. Instance object detection is one of mean service that robots needs. Classical approaches are based on extracting discriminant and invariant features. However those features still have a limitation to represent all kinds of objects and satisfy all requirements (discrimination and invariance). Since no single feature can work well in various situations, we need to combine several features so that the robot can handle all kind of daily life objects. Our task consists in defining a strategy that can work on various objects and backgrounds without any prior knowledge. In this paper we propose a scheme to combine two descriptors using belief function theory. First, objects are extracted from image and described by two complementary descriptors: Dominant Color Descriptor for color description and Zernike Moments for shape description. Second, similarity indicators is computed between object of interest descriptors and each extracted object descriptors. Finally, those measures are combined into a belief functions in order to build a final decision about the object presence in the image taking the information uncertainty and imprecise into consideration. We have evaluated our approach with different methods of information fusion such as the weighted vote approach, the possibility theory and so forth.

In this paper we present a new method for key frame detection. Our approach is based on a well known algorithm: Speeded-Up Robust Features (SURF), which is a crucial step of our method. The frames are compared by a SURF matcher, which allows to count the corresponding keypoints. The proposed method provides better results for professional and high resolution videos. The simulations we conducted proved the effectiveness of our approach. The algorithm requires only one input parameter.

Dermatology is one of the fields where computer aided diagnostic is developing rapidly. The presented research concentrates on creation of automatic methods for melanoid skin lesions diagnosis using machine learning methods. In the experiments 1010 samples described in [5] are used. There are 275 melanoma cases and 735 benign ones. Three different machine learning methods are applied, namely the Naive Bayes classifier, the Random Forest, the K* instance-based classifier, and Attributional Calculus. The obtained results confirm that clinical history context and dermoscopic structures together with the selected machine learning methods may be an important and accurate diagnostic tool.

In this paper, we propose an iterative algorithm for edge detection. The presented method uses a 2D Gaussian function and Prewitt operator. Firstly, the selection of edge points has been achieved by a statistical method. Secondly, the boundary tracking has been performed by the rotation of the Gaussian function and by changing function’s variables. The algorithm has been tested using a medical phantom. Additionally, an implementation on multicore GPUs has been designed for a better performance.

The possibility of correct automatic recognition of emotion is of high importance in many research domains. In this paper the choice of relevant face regions for smile detection is investigated. Firstly, three facial display division schemas are compared. Afterwards, for the most promising, some region masks are suggested. The presented approach differs in feature vector length. The performance of classification between smiling and neutral facial display proved that applying masks, hence shortening the feature vector, does not decrease the accuracy but even improves the result.

The Liquid Argon Time Projection Chamber (LAr-TPC) detectors provide excellent imaging and particle identification ability for studying neutrinos. An efficient and automatic reconstruction procedures are required to exploit potential of this imaging technology. Herein, a novel method for segmentation of images from LAr-TPC detectors is presented. The proposed approach computes a feature descriptor for each pixel in the image, which characterizes amplitude distribution in pixel and its neighbourhood. The supervised classifier is employed to distinguish between pixels representing particle’s track and noise. The classifier is trained and evaluated on the hand-labeled dataset. The proposed approach can be a preprocessing step for reconstructing algorithms working directly on detector images.

Our aim in this paper is to extend the results on parallel change detection recently discussed in [15], where such a detector has been proposed. Here, we emphasize its adaptive abilities to follow changing background and relax some theoretical assumptions on random errors, extending possible applications of the detector. We also discuss its implementation in NVidia CUDA technology and provide results of its extensive testing when applied to copper visual quality control, which is a challenge due to the need for massively parallel calculations in real-time.

In this paper we have proposed a gender recognition system through facial images. We have used three different techniques that involve Bandlet Trans-form (a multi-resolution technique), LBP (Local Binary Pattern) and mean to create the feature vectors of the images. To classify the images for gender, we have used fuzzy c mean clustering. SUMS and FERET databases were used for testing. Experimental results have shown that the maximum average accuracy was achieved using SUMS, 97.1% has been achieved using Band-lets and mean technique, Bandlets and whole image LBP has shown 85.13% and Bandlets with blocked based LBP has shown 87.02% average accuracy.

The principle of justifiable granularity, as formulated in [1], defines intuitively motivated requirements for an information granule to be meaningful. In the paper, granulation of images obtained by their segmentation is considered. In this context, such concepts as representation of granules and their relations, representation of concepts, consideration of context, detection and treatment of outliers, and recognition method, are of importance. The granular approach is related to intelligent analysis of all kinds of data, not only the computer images.

The paper concerns specific problems of color digital image recognition by use of the concept of fuzzy and rough granulation. This idea employs information granules that contain pieces of knowledge about digital pictures such as color, location, size, and shape of an object to be recognized. The object information granule (OIG) is introduced, and the Granular Pattern Recognition System (GPRS) proposed, in order to solve different tasks formulated with regard to the information granules.

This paper is to discuss a matter of preprocessing 2D input images by selected methods of Evolutionary Computation. In the following sections we try to analyze possibility of using Artificial Bee Colony algorithm to preprocess input images for classification purposes. Experiments have been performed with the examined method applied on a set of test images, to present and discuss efficacy and precision of recognition.

The paper presents our approach to SVM implementation in parallel environment. We describe how classification learning and prediction phases were pararellised. We also propose a method for limiting the number of necessary computations during classifier construction. Our method, named

one-vs-near

, is an extension of typical

one-vs-all

approach that is used for binary classifiers to work with multiclass problems. We perform experiments of scalability and quality of the implementation. The results show that the proposed solution allows to scale up SVM that gives reasonable quality results. The proposed

one-vs-near

method significantly improves effectiveness of the classifier construction.

The standard SVM classifier is not adjusted to processing large training set as the computational complexity can reach

O

(

n

3

). To overcome this limitation we discuss the idea of reducing the size of the training data by initial preprocessing of the training set using Learning Vector Quantization (LVQ) neural network and then building the SVM model using prototypes returned by the LVQ network. As the LVQ network scales linearly with

n

, and in contrast to clustering algorithms utilizes label information it seems to be a good choice for initial data compression.

We analyze the privacy preservation capabilities of a previously introduced multi-stage image representation framework where blocks of images with similar statistics are decomposed into different codebooks (dictionaries). There it was shown that at very low rate regimes, the method is capable of compressing images that come from the same family with results superior to those of the JPEG2000 codec. We consider two different elements to be added to the discussed approach to achieve a joint compression-encryption framework. The first visual scrambling is the random projections were the random matrix is kept secret between the encryption and decryption sides. We show that for the second approach, scrambling in the DCT domain, we can even slightly increase the compression performance of the multi-layer approach while making it safe against de-scrambling attacks. The experiments were carried out on the

ExtendedYaleB

database of facial images.

Growing Self Organized Map (GSOM) algorithm is a well-known unsupervised clustering algorithm which a definite advantage is that both the map structure as well as the number of classes are automatically adjusted depending on the training data. We propose a new approach to apply it in the process of the image indexation and retrieval in a database. Unlike the classic bag-of-words (BoW) algorithm with

k

-means clustering, it is completely unnecessary to predetermine the number of classes (words). Thanks to that, the process of indexation can be fully automated. What is more, numerous modifications of the classic algorithm were added, and as a result, the retrieval process was considerably improved. Results of the experiments as well as comparison with BoW are presented at the end of the paper.

In this paper we present a novel architecture for storing visual data. Effective storing, browsing and searching collections of images is one of the most important challenges of computer science. The design of architecture for storing such data requires a set of tools and frameworks such as SQL database management systems and service-oriented frameworks. The proposed solution is based on a multi-layer architecture, which allows to replace any component without recompilation of other components. The approach contains five components, i.e. Model, Base Engine, Concrete Engine, CBIR service and Presentation. They were based on two well-known design patterns: Dependency Injection and Inverse of Control. For experimental purposes we implemented the SURF local interest point detector as a feature extractor and

K

-means clustering as indexer. The presented architecture is intended for content-based retrieval systems simulation purposes as well as for real-world CBIR tasks.

In this work a method of object localization on the basis of its expected structure is presented. An active partition approach is used for that purpose where, instead of pixels, line segments are used to represent image content. The expectation about object being sought is expressed in the form of model where the expected line segments are specified explicitly. Both image representation and model take into account relations between segments and thus both can be considered as graphs constituting their structural description. The best subsets of line segments are sought in a systematic search process with properly defined model fit function. It allows to identify a subset of segments that resembles the given model even if the segments are detected imprecisely.

In this paper we investigate transform learning and apply it to face recognition problem. The focus is to find a transformation matrix that transforms the signal into a robust to noise, discriminative and compact representation. We propose a method that finds an optimal transform under the above constrains. The non-sparse variant of the presented method has a closed form solution whereas the sparse one may be formulated as a solution to a sparsity regularized problem. In addition we give a generalized version of the proposed problem and we propose a prior on the data distribution across the dimensions in the transform domain.

Supervised transform learning is applied to the MVQ [10] method and is tested on a face recognition application using the YALE B database. The recognition rate and the robustness to noise is superior compared to the original MVQ based on

k

-means.

This paper describes a method for searching for common sets of descriptors between collections of images. The presented method operates on local interest keypoints, which are generated using the SURF algorithm. The use of a dictionary of descriptors allowed achieving good performance of the content-based image retrieval. The method can be used to initially determine a set of similar pairs of keypoints between images. For this purpose, we use a certain level of tolerance between values of descriptors, as values of feature descriptors are almost never equal but similar between different images. After that, the method compares the structure of rotation and location of interest points in one image with the point structure in other images. Thus, we were able to find similar areas in images and determine the level of similarity between them, even when images contain different scenes.

According to reports from medical community the number of autistic children’s birth is more and more alarming. Early diagnosis and regular rehabilitation are crucial. The problem with verbal and emotional communication is very common. In a form of short survey, a few similar issues and their solutions have been examined in terms of input data type, feature selection, pattern recognition and formal mathematical modeling. Then we propose a system for autistic children rehabilitation, surveillance and emotions translation. These new solutions have been compared with those reported in the literature. The preliminary experiments provide rather satisfactory results.

In video surveillance, we can rely on either a visible spectrum or an infrared one. In order to profit from both of them, several fusion methods were proposed in literature: low-level fusion, middle-level fusion and high-level fusion. The first one is the most used for moving objects’ detection. It consists in merging information from visible image and infrared one into a new synthetic image to detect objects. However, the fusion process may not preserve all relevant information. In addition, perfect correlation between the two spectrums is needed. In This paper, we propose an intelligent fusion method for moving object detection. The proposed method relies on one of the two given spectrum at once according to weather conditions (darkness, sunny days, fog, snow, etc.). Thus, we first extract a set of low-level features (visibility, local contrast, sharpness, hue, saturation and value), then a prediction model is generated by supervised learning techniques. The classification results on 15 sequences with different weather conditions indicate the effectiveness of the extracted features, by using C4.5 as classifier.

This paper describes a method for rapid location and characterization of objects in 2D images. It derives optimizing parameters of a normalized Gaussian that best approximates the observed object, simultaneously finding the object location in the observed scene. A similarity measure to this optimized Gaussian is used to characterize the object. Optimization process has global and exponentially fast convergence, thus it can be used to implement saccadic motion for object recognition and scene analysis. This method was inspired by Perlovsky’s work on neural dynamic logic used for fast location, characterization, and identification of objects. Developed method was tested and illustrated with an example of an object location and characterization.

Writing user-defined functions or stored procedures presents common way in application development using a relational database management system. It allows to embed application code inside of RDBMS. In this paper, we examine an effect of embedding selected computer vision algorithms as user-defined functions in a relational database management system. We show that such a combination can in certain scenarios lead to a performance improvement.