Artificial Intelligence and Soft Computing

We consider a scenario where an automatic classifier has been built, but it sometimes decides to ask the correct label of an instance to an oracle, instead of accepting its own prediction. This interactive classifier only knows with certainty the labels provided by the oracle. Our proposal is to use this information to dynamically improve the behavior of the classifier, either increasing its accuracy when it is being used autonomously or reducing the number of queries to the oracle. We have tested our proposal by using twenty data sets and two adaptive classifiers from the Massive Online Analysis (MOA) open source framework for data stream mining.

Clustering is one of the major tasks in data mining. However, selecting an algorithm to cluster a dataset is a difficult task, especially if there is no prior knowledge on the structure of the data. Consensus clustering methods can be used to combine multiple base clusterings into a new solution that provides better partitioning. In this work, we present a new consensus clustering method based on detecting clustering patterns by mining frequent closed itemset. Instead of generating one consensus, this method both generates multiple consensuses based on varying the number of base clusterings, and links these solutions in a hierarchical representation that eases the selection of the best clustering. This hierarchical view also provides an analysis tool, for example to discover strong clusters or outlier instances.

In this paper we present experimental results on rule sets induced from 12 data sets with many missing attribute values. We use two interpretations of missing attribute values: lost values and attribute-concept values. Our main objective is to check which interpretation of missing attribute values is better from the view point of complexity of rule sets induced from the data sets with many missing attribute values. The better interpretation is the attribute-value. Our secondary objective is to test which of the three probabilistic approximations used for the experiments provide the simplest rule sets: singleton, subset or concept. The subset probabilistic approximation is the best, with 5 % significance level.

In this paper an incremental procedure for nonparametric learning of time-varying regression function is presented. The procedure is based on the Cesàro-means of orthogonal series. Its tracking properties are investigated and convergence in probability is shown. Numerical simulations are performed using the Fejer’s kernels of the Fourier orthogonal series.

In this article we investigate the problem of regression functions estimation in the edges points of their domain. We refer to the model $$y_i = R\left( {x_i } \right) + \epsilon _i ,\,i = 1,2, \ldots n$$yi=Rxi+ϵi,i=1,2,…n, where $$x_i$$xi is assumed to be the set of deterministic inputs, $$x_i \in D$$xi∈D, $$y_i$$yi is the set of probabilistic outputs, and $$\epsilon _i$$ϵi is a measurement noise with zero mean and bounded variance. R(.) is a completely unknown function. The possible solution of finding unknown function is to apply the algorithms based on the Parzen kernel [13, 31]. The commonly known drawback of these algorithms is that the error of estimation dramatically increases if the point of estimation x is drifting to the left or right bound of interval D. This fact makes it impossible to estimate functions exactly in edge values of domain.The main goal of this paper is an application of NMS algorithm (introduced in [11]), basing on integral version of the Parzen method of function estimation by combining the linear approximation idea. The results of numerical experiments are presented.

In this paper the issue of splitting criteria used in decision tree induction algorithm designed for data streams is analyzed. A hybrid splitting criterion is proposed which combines two criteria established for two different split measure functions: the Gini gain and the split measure based on the misclassification error. The hybrid splitting criterion reveals advantages of its both component. The online decision tree with hybrid criterion demonstrates higher classification accuracy than the online decision trees with both considered single criteria.

In the paper a problem of outlier detection in the stream data is raised. The authors propose a new approach, using well known outlier detection algorithms, of outlier detection in the stream data. The method is based on the definition of a sliding window, which means a sequence of stream data observations from the past that are closest to the newly coming object. As it may be expected the outlier detection accuracy level of this model becomes worse than the accuracy of the model that uses all historical data, but from the statistical point of view the difference is not significant. In the paper several well known methods of outlier detection are used as the basis of the model.

Classification of solar event detections into two classes, of either the same object at a later time or an entirely different object, plays a significant role in multiple hypothesis solar event tracking. Many features for this task are produced when images from multiple wavelengths are used and compounded when multiple image parameters are extracted from each of these observations coming from NASA’s Solar Dynamics Observatory. Furthermore, each different event type may require different sets of features to accurately accomplish this task. A feature selection algorithm is required to identify important features extracted from the available images and that can do so without a high computational cost. This work investigates the use of a simple feature subset selection method based on the ANOVA F-Statistic measure as a means of ranking the extracted image parameters in various wavelengths. We show that the feature subsets that are obtained through selecting the top K features ranked in this manner produce classification results as good or better than more complicated methods based on searching the feature subset space for maximum-relevance and minimum-redundancy. We intend for the results of this work to lay the foundations of future work towards a robust model of appearance to be used in the tracking of solar phenomena.

Text data mining is the process of extracting valuable information from a dataset consisting of text documents. Popular clustering algorithms do not allow detection of the same words appearing in multiple documents. Instead, they discover general similarity of such documents. This article presents the application of a hybrid biclustering algorithm for text mining documents collected from Twitter and symbolic analysis of knowledge spreadsheets. The proposed method automatically reveals words appearing together in multiple texts. The proposed approach is compared to some of the most recognized clustering algorithms and shows the advantage of biclustering over clustering in text mining. Finally, the method is confronted with other biclustering methods in the task of classification.

In this paper a modification of the well-known Silhouette validity index is proposed. This index, which can be considered a measure of the data set partitioning accuracy, enjoys significant popularity and is often used by researchers. The proposed modification involves using an additional component in the original index. This approach improves performance of the index and provides better results during a clustering process, especially when changes of cluster separability are big. The new version of the index is called the SILA index and its maximum value identifies the best clustering scheme. The performance of the new index is demonstrated for several data sets, where the popular algorithm has been applied as underlying clustering techniques, namely the Complete–linkage algorithm. The results prove superiority of the new approach as compared to the original Silhouette validity index.

Data dissimilarities and similarities are the key ingredients of machine learning. We give a mathematical characterization and classification of those measures based on structural properties also involving psychological-cognitive aspects of similarity determination, and investigate admissible conversions. Finally, we discuss some consequences of the obtained taxonomy and their implications for machine learning algorithms.

Monitoring of human behavior in the natural living habitat requires a hidden yet accurate measurement. Several previous attempts showed, that this can be achieved by recording and analysing interactions of the supervised human with sensorized equipment of his or her household. We propose an imperceptible single-sensor measurement, already applied for energy usage profiling, to detect the usage of electrically powered domestic appliances and deduct important facts about the operator’s functional health. This paper proposes a general scheme of the system, discusses the personalization and adaptation issues and reveals benefits and limitations of the proposed approach. It also presents experimental results showing reliability of device detection based on their load signatures and areas of applicability of the load sensor to analyses of device usage and human performance.

The main purpose of this paper is to present the properties of our novel statistical model-based iterative approach to the image reconstruction from projections problem regarding its condition number. The reconstruction algorithm based on this concept uses a maximum likelihood estimation with an objective adjusted to the probability distribution of measured signals obtained using x-ray computed tomography. We compare this with some selected methods of regularizing the problem. The concept presented here is fundamental for 3D statistical tailored reconstruction methods designed for x-ray computed tomography.

The nosocomial infections are a growing concern because they affect a large number of people and they increase the admission time in healthcare facilities. Additionally, its diagnosis is very tricky, requiring multiple medical exams. So, this work is focused on the development of a clinical decision support system to prevent these events from happening. The proposed solution is unique once it caters for the explicit treatment of incomplete, unknown, or even contradictory information under a logic programming basis, that to our knowledge is something that happens for the first time.

The increasing incidence of melanoma skin cancer is alarming. The lack of objective diagnostic procedures encourages development of computer aided approaches. Presented research uses three different machine learning methods, namely the Naive Bayes classifier, the Random Forest and the K* instance-based classifier together with two meta-learning algorithms: the Bootstrap Aggregating (Bagging) and the Vote Ensemble Classifier. Diagnostic accuracy of the selected methods, such as sensitivity and specificity and the area under the ROC curve, are discussed. The obtained results confirm that clinical history context and dermoscopic structures present in the images are important and can give accurate diagnostic classification of the lesions.

An approach permitting to build free schedules for the RNA folding algorithm is proposed. The statements can be executed in parallel as soon as all their operands are available. This technique requires exact dependence analysis for automatic parallelization of the Nussinov algorithm. To describe and implement the algorithm the dependence analysis by Pugh and Wonnacott was chosen where dependencies are found in the form of tuple relations. The approach has been implemented and verified by means of the islpy and CLooG tools as a part of the TRACO compiler. The experimental study presents speed-up, scalability and costs of parallelism of the output code. Related work and future tasks are described.

Since the number of autistic children births increases each year, Autism Spectrum Disorder has become a serious community problem. In this paper we present the development of an integrated system for children with autism (surveillance, rehabilitation and daily life assistance). The hierarchical classifier for human position recognition has been developed and the scalable symbols codebook for Hidden Markov Models has been created. For data acquisition Microsoft Kinect 2.0 depth sensor is used. A few experiments for basic action models have been conducted and the preliminary results are satisfactory. The obtained classifiers will be used in further work.

The purpose of this article is to discuss the motivation and benefits of developing and releasing a cloud service providing digital signature in software and infrastructure as-a-service model. Additionally, since users authorization and authentication is based on biometry (analyzing the blood vessels system) the end user doesnt have to be equipped with any additional smart-cards or devices for storing the private key and performing crypto-operations, and the only what he needs to digitally sign data, files or documents is a web browser and his finger. Podpiszpalcem.pl is the service realizing the above idea and is presented in this paper.

Landmarks are unique points that can be located on every face. Facial landmarks typically recognized by people are correlated with anthropomorphic points. Our purpose is to employ in 3D face recognition such landmarks that are easy to interpret. Face understanding is construed as identification of face characteristic points with automatic labeling of them. In this paper, we apply methods based on Self Organizing Maps to understand 3D faces.

Identity verification using the dynamic signature is an important biometric issue. Its big advantage is that it is commonly socially acceptable. Verification based on so-called global features is one of the most effective methods used for this purpose. In this paper we propose an approach which minimises a number of the features used during verification process due to check how the number of features affects the classification result. The paper contains the simulation results for the public MCYT-100 database of the dynamic signatures.

Dynamic signature verification is a very interesting biometric issue. It is difficult to realize because signatures of the user are characterized by relatively high intra-class and low inter-class variability. However, this method of an identity verification is commonly socially acceptable. It is a big advantage of the dynamic signature biometric attribute. In this paper we propose a new hybrid algorithm for the dynamic signature verification based on global and regional approach. We present the simulation results of the proposed method for BioSecure DS2 database, distributed by the BioSecure Association.

The purpose of nonlinear correction modelling of dynamic object is to use an approximated linear model of an object and determine corrections of this model in an appropriate way, taking into account the specificity of modelled nonlinearity. In this paper a new method for generating the coefficients of correction matrices is proposed. This method uses a mathematical formulas determined automatically by the Gene Expression Programming algorithm extended by semantic operator.

In this paper we propose a new approach for nonlinear modelling. It uses capabilities of the Takagi-Sugeno neuro-fuzzy systems and population based algorithms. The aim of our method is to ensure that created model achieves appropriate accuracy and is as compact as possible. In order to obtain this aim we incorporate semantic information about created fuzzy rules into process of evolution. Our method was tested with the use of well-known benchmarks from the literature.

The paper presents a new approach for using the fuzzy decision tress for the detection of the significant operating points from non-invasive measurements of the nonlinear dynamic object. The PSO-GA algorithm is used to identify the unknown values of the system matrix describing the nonlinear dynamic object. It is defined in the terminal nodes of the fuzzy decision tree. The new approach was tested on the nonlinear electrical circuit. The obtained results prove efficiency of the new approach for using fuzzy decision tree for the detection of the significant operating points in the nonlinear modeling.

The paper presents a new method of the intelligent modeling of the nonlinear dynamic objects with online detection of significant operating points from non-invasive measurements of the nonlinear dynamic object. The PSO-GA algorithm is used to identify the unknown values of the system matrix describing the nonlinear dynamic object in the detected operating points. The Takagi-Sugeno fuzzy system determines the values of the system matrix in the detected operating points. The new method was tested on the nonlinear electrical circuit with the three operating points. The obtained results prove efficiency of the new method of the intelligent modeling of the nonlinear dynamic objects with fuzzy detection of the operating points.

This paper is aimed at the evaluating of utility of 3 temporal logics: linear temporal logic (LTL) and Halpern-Shoham interval logic from the point of view of the engineering practice. We intend to defend the thesis that chosen systems are only partially capable of satisfying typical requirements of engineers.

This paper is aimed at proposing some new formal system of a fuzzy logic – suitable for representation the “before” relation between temporal intervals. This system and an idea of the integral-based approach to the representation of the Allen’s relations between temporal intervals is later used for a specification of a class of solutions of the so-called Simple Temporal Problem under Uncertainty and it extends the classical considerations of R. Dechter and L. Khatib in this area.

Formal dialogue systems are an important trend in current research on the process of communication. They can be used as the schema of the dialogue conducted between artificial entities or as a simplified form of human dialogue with a machine or a human being with a man. In this work we introduce a mathematical model of dialogue, which is inspired by dialogue games. This model will be used as a semantic structure in verification of properties of dialogue protocols. For this purpose, the semantics of the dialogue games has been translated into interpreted systems that are commonly used in the model checking approach. The newly created model will be applied to develop methods and techniques for automated analysis of dialogues.

Time-dependent behavior of queueing delay distribution in the GI/M/1/K-type model with the “queued” server’s waking up and balking is studied. After each idle period the server is being “queued” woken up, i.e. the processing is being started at the moment the number of packets accumulated in the buffer reaches the fixed level N. Moreover, each incoming packet can balk (resign from service) and leave the system irrevocably, with probability $$1-\beta ,$$ and join the queue with probability $$\beta ,$$ where $$0< \beta \le 1$$.

The collective knowledge increase, in general, is understood as an additional amount of knowledge in a collective in comparison with the average of the knowledge states given by collective members on the same subject in the real world. These knowledge states reflect the real knowledge state of the subject, but only to some degree because of the incompleteness and uncertainty. In this work, we investigate the influence of the inconsistency degree on the collective knowledge increase in a collective by taking into account the number of collective members. In addition, by means of experiments we prove that the amount of knowledge increase in a collective with higher inconsistency degree is better than that in a collective with lower inconsistency degree.

This paper is a continuation and extension in developing the knowledge-based decision support design system (called HSSDR) which communicates with the designer via drawings. Graph-based modeling of conceptualization in the CAD process, which enables the system to automatically transform design drawings into appropriate graph-based data structures, is considered. Hierarchical graphs with bonds are proposed as a representation of designs. An ontological commitment between design conceptualization and internal representations of solutions, which enables us to capture intended design models, is described. Moreover, the first-order logic (FOL) of HSSDR is replaced by many-sorted FOL that makes it possible to define different sorts in specification of functions and predicates in semantics and design constraint verification.

The paper presents a new version on the mini-models method (MM-method). Generally, the MM-method identifies not the full global model of a system but only a local model of the neighborhood of the query point of our special interest. It is an instance-based learning method similarly as the k-nearest algorithm, GRNN network or RBF network but its idea is different. In the MM-method the learning process is based on a group of points that is constrained by a polytope. The first MM-method was described in previous publications of authors. In this paper a new version of the MM-method is presented. In comparison to the previous version it was extended by local dimensionality reduction. As experiments have shown this reduction not only simplifies local models but also in most cases allows for increasing the local model precision.

One of exploitation methods of liquid fossil fuel deposits depends on pumping chemicals to the geological formation and ‘sucking out’ the fuel that is pushed out by the solution. This method became particularly popular in the case of extraction of shale gases. A real problem here is however a natural environment contamination caused mainly by chemicals soaking through the geological formations to ground-waters.The process of pumping the chemical fluid into the formation and extracting the oil/gas is modeled here as a non-stationary flow of the non-linear fluid in heterogeneous media.The (poly)optimization problem of extracting oil in such a process is then defined as a multiobjetcive optimization problem with two contradictory objectives: maximizing the amount of the oil/gas extracted and minimizing the contamination of the ground-waters.To solve the problem defined a hibridized solver of multiobjective optimization of liquid fossil fuel extraction (LFFEP) integrating population-based heuristic (i.e. NSGA-II algorithm for approaching the Pareto frontier) with isogeometric finite element method IGA-FEM method for modeling non-stationary flow of the non-linear fluid in heterogeneous media is presented along with some preliminary experimental results.

This participation is focused on artificial intelligence techniques and their practical use in computer game. The aim is to show how game player (based on evolutionary algorithms) can replace a man in two computer games. The first one is strategy game StarCraft: Brood War, briefly reported here. Implementation used in our experiments use classic techniques of artificial intelligence environments, as well as unconventional techniques, such as evolutionary computation. The second game is Tic-Tac-Toe in which SOMA has also take a role of player against human player. This provides an opportunity for effective, coordinated movement in the game fitness landscape. Research reported here has shown potential benefit of evolutionary computation in the field of strategy games and players strategy mining based on their mutual interactions.

In the work there is a NP-hard cyclic job shop problem of tasks scheduling considered. To its solution there was tabu search algorithm implemented using neural mechanism to prevent looping of the algorithm. There were computational experiments conducted that showed statistically significant efficacy of the proposed tabu method as compared to classical list of forbidden moves.

In many real production systems the parameters of individual operations are not deterministic. Typically, they can be modeled by fuzzy numbers or distributions of random variables. In this paper we consider the flexible job shop problem with machine setups and uncertain times of operation execution. Not only we present parallel algorithm on GPU with fuzzy parameters but also we investigate its resistance to random disturbance of the input data.

We present a new method of quantitative graph analysis and visualization based on vertex centrality measures and distance matrices. After generating distance k-graphs and collecting frequency information about their vertex descriptors, we obtain generic, multidimensional representation of a graph, invariant to graph isomorphism. The histograms of vertex centrality measures, organized in a form of B-matrices, allow to capture subtle changes in network structure during its evolution and provide robust tool for graph comparison and classification. We show that different types of B-matrices and their extensions are useful in graph analysis tasks performed on benchmark complex networks from Koblenz and IAM datasets. We compare the results obtained for proposed B-matrix extensions with performance of other state-of-art graph descriptors showing that our method is superior to others.

Production systems use forward chaining to perform the reasoning, in this case - matching rules with facts. The Rete algorithm is an effective forward chaining realization. With the growing popularity of functional programming style, questions arise, how well suitable the style is for implementing complex algorithms in the Artificial Intelligence, like Rete. We present selected implementation details of our custom realization of the algorithm in purely functional programming language Haskell. This paper also discusses usability and usefulness of some advanced means of expression, that are common in functional style, for performing the task.

The same business concepts can be expressed in various knowledge representations like processes or rules. This paper presents an interoperability solution for transforming a subset of the SBVR rules into the BPMN and DMN models. The translation algorithm describes how to translate the SBVR vocabulary, structural and operational rules into particular BPMN and DMN elements. The result is a combined process and decision model, which can be used for validating SBVR rules by people aware of BPMN and DMN notations.

Graphs are an expressive representation of projects in the domain of computer-aided design (CAD). Such a representation of a problem’s structure allows for automation of a design process, what is an important property of CAD systems. It can be accomplished by using graph grammars which can represent a progress of a design process. In this paper we introduce a graph based approach to the synchronization of a design processes carried out by different and independent transformation systems supporting various aspects of a building project creation. Such a synchronization is necessary when two or more systems affect simultaneously a shared area. The proposed mechanism is illustrated by an example of successful synchronization on shared elements of an object being designed, achieved by using different representations at different layers of a design.

This paper examines the machine learning approach to authorship attribution of articles in the Polish language. The focus is on the effect of the data volume, number of authors and thematic homogeneity on authorship attribution quality. We study the impact of feature selection under various feature selection criteria, mainly chi square and information gain measures, as well as the effect of combining features of different types. Results are reported for the Rzeczpospolita corpus in terms of the $$F_1$$F1 measure.

This paper presents a new adaptation of the cockroach swarm optimization (CSO) algorithm to effectively solve the traveling salesman problem. Proposed modifications investigate the crossover operators in chase-swarming procedure and directed dispersion of cockroaches. To analyze the benefits of such modifications, the performance of the considered approach is tested on well-known instances. Presented results of all experiments indicate that practical implementation of the CSO algorithm, which includes the sequential constructive crossover (SCX) and 2-opt move is a good approach.

The paper concerns the main element of the molecular neural network - the Molecular Neuron (MN). Molecular Neural Network idea has been introduced in our previous articles. Here we present the structure of the Molecular Neuron element in micro and nanoscale. We have obtained MN in hexagonal layout in the form of the thin film. In this paper we have described self-assembly mechanism leading to the NMs layout. Also physical properties of the MNs layer have been shown.

Crowdsourcing platforms are very frequently used for collecting training data. Quality assurance is the most obvious problem but not the only one. This work proposes iterative approach which helps to reduce costs of building training/testing datasets. Information about classifier confidence is used for making decision whether new labels from crowdsourcing platform are required for this particular object. Conducted experiments have confirmed that proposed method reduces costs by over 50 % in best scenarios and at the same time increases the percentage of correctly classified objects.

This article presents a method of localizing a moving mobile terminal (i.e. phone) with the usage of the Particle Filter method. The method is additionally enhanced with the predictions done by a Random Forest and the results are optimized with the usage of the Particle Swarm Optimization algorithm.The method proposes a simple model of movement through the building, a likelihood estimation function for evaluating locations against the observed signal, and a method of generating multiple location propositions from a single point prediction statistical model on the basis of model error estimation.The method uses a data set of the GSM and WiFi networks received signals’ strengths labeled with a receiver’s 3D location. The data have been gathered in a six floor building. The approach is tested on a real-world data set and compared with a single point estimation performed by a Random Forest. The Particle Filter approach has been able to improve floor recognition accuracy by around $$7\,\%$$7% and lower the median of the horizontal location error by around $$15\,\%$$15%.

In this paper the potentials of identifying unusual user behaviors and changes of behavior from smart home energy meters are investigated. We compare the performance of the classical change detection Page-Hinkley test (PHT) with a new application of a self-adaptive stream clustering algorithm to detect novelties related to the time of use of appliances at home. With the use of annotated data, the true positive rate of the clustering-based method outperformed the PHT by at least 20 %. Moreover the method was able to identify behavior changes related to time shifts and replacement of appliances. The motivation for this study is based on the need for identifying and guiding behavior changes that can reduce energy consumption, and use this knowledge in the development of systems that can raise just-in-time warnings to save energy (e.g. avoid stand-by modes), and guide sustainable behavior changes.

Nanotechnology is based on molecules with spin energy [4, 7, 14]. These elementary particles are located outside magnetic field [5, 15, 17]. Due to their chemical structure they react differently in reference to their own magnetic spin value correction [18]. The correction refers also to the spin direction (phase $$\theta $$θ) [20]. Technology parameters and conditions will not be considered in the paper. We focus on idea of finding the most correlated memory module location with given key structure. The problem is difficult as existing solutions in traditional computer technology with memory cells fitting comparator do not conform to the spin technology set of tools. Nevertheless, when we implement operation based on probabilities of binary states after each operation we lose measured value and phase [8]. Therefore, our proposition should be based on different strategy of finding the closest distance among key and context of memory blocks.

The paper proposes a new class of the hardware emulators, namely the remote emulators. They can temporarily replace a control object to allow testing of a distributed system in a safe manner. This method is named a remote-hardware-in-the-loop (RHIL). The second issue described in the paper is a hybrid method of using the computational intelligence in the hardware emulators. This hybrid system is based on a radial-basis-function, a fuzzy-logic and a state variables theory. The proposed solutions make it possible to build a hardware emulator that can work in the RHIL systems with a good accuracy.

Transport conditions have a direct impact on the costs of supply in the distribution network. In the wide area networks differences in transport costs depending on external conditions achieve significant meaning. The issue concerns not only the wheel transportation but also its other forms, including the transmission of different types of media (flow). However, it is felt most clearly in case of delivery with the use of the vehicle base. Among the conditions affecting the costs, one can distinguish internal (type of vehicle, the loading size) and external factors (variable capacity resulting for example from traffic, various forms of traffic disturbances). This leads to the conclusion, that costs can and should be estimated in the interval form. The consequence of such analysis will be choosing the cheapest connection configurations which will be supplemented by the system of the inference rules [11]. Such an approach is presented in the rough sets theory.

The paper concerns bank marketing selling campaign result prediction by neuro-fuzzy systems. We trained the system by the backpropagation algorithm using forty five thousand of past records. We obtained comparable prediction results with the best ones from the literature. The advantage of the proposed approach is the use of fuzzy rules, which are interpretable for humans.

The confidence intervals in computer Go are used in MCTS algorithm to select the potentially most promising moves that should be evaluated with Monte-Carlo simulations. Smart selection of moves for evaluation has the crucial impact on program’s playing strength. This paper describes the application of confidence intervals for binomial distributed random variables in computer Go. In practice, the estimation of confidence intervals of binomial distribution is difficult and computationally exhausted. Now due to computer technology progress and functions offered by many libraries calculation of confidence intervals for discreet, binomial distribution become an easy task. This research shows that the move-selection strategy which implements calculation of the exact confidence intervals based on discreet, binomial distribution is much more effective than based on normal. The new approach shows its advantages particularly in games played on medium and large boards.

We introduce a texture-based modification of the GrabCut algorithm that significantly improves its performance for high-resolution images but with a slight decrease in accuracy. This consists of five steps: expansion, convolution, shrinkage, GrabCut of the shrunk image, and enlargement. The results showed that modified algorithm is three times faster than the original one. At the same time, there is no significant difference between the average F1-measures obtained for both algorithms in case of high-resolution images. Therefore, it can be successfully used in semi-automatic segmentation of such images.

This work describes a set of techniques, based on the vision system, designed to supplement information about environment by adding three-dimensional objects representations. Described vision system plays a role of supplementary part of the SLAM technique for gathering information about surrounding environment by an autonomous robot. Algorithms are especially prepared for a mobile drilling robot. The main characteristics of the robot and its applications are defined in the first part of this paper. Then, the technical aspects and the execution steps of the algorithms utilized by the vision system are described. In the last part of this paper, the test case along with the results, presenting sample application of the vision system, is presented.

In this work a new software system and environment for detecting objects with specific features within an image is presented. The developed system has been applied to a set of satellite transmitted SAR images, for the purpose of identifying objects like ships with their wake and oil slicks. The systems most interesting characteristic is its flexibility and adaptability to largely different classes of objects and images, which are of interest for several application areas. The heart of the system is represented by the clustering subsystem. This is to extract from the image objects characterized by local properties of small pixel neighborhoods. Among these objects the desired one is sought in later stages by a classifier to be plugged in, chosen from a pool including both soft-computing and conventional ones. An example of application of the system to a recognition problem is presented. The application task is to identify objects like ships with their wake and oil slicks within a set of satellite transmitted SAR images. The reported results have been obtained using a back-propagation neural network.

This article summarizes the achievements that have been made in the field of GPU SVM acceleration. In particular, the algorithms which allow the acceleration of SVM classification performed on dense datasets are presented and the limitations of the dataset size are pointed out. Moreover, the solutions which deal with large sparse collections are demonstrated. These algorithms apply different sparse dataset formats to make possible the classification on the GPU. Finally, GPU implementations for different SVM kernel functions are provided.

This article presents a modification of the Bag-of-Features method (also known as a Bag-of-Words or Bag-of-Visual-Words method) used for image recognition in practical applications using a relational database. Our approach utilises a modified k-means algorithm, owing to which the number of clusters is automatically selected, and also the majority votes method when making decisions in the classification process. The algorithm can be used both methods in an SQL Server database or a commonly-used MySQL one. The proposed approach minimises the necessity to use additional algorithms and/or classifiers in the image classification process. This makes it possible to significantly simplify computations and use the SQL language.

In this paper we present a novel approach to image description. Our method is based on the Canny edge detection. After the edge detection process we apply a self-designed crawler method. The presented algorithm uses edges in order to move on pixel edges and describe the entire object. Our approach is closely related with the content-based image retrieval and it can be used as a pre-processing stage but can also be used for general purpose image description. The experiments proved the effectiveness of our method as it provides better results then the SURF descriptor.

Video and image compression technology has evolved into a highly developed field of computer vision. It is used in a wide range of applications like HDTV, video transmission, and broadcast digital video. In this paper the new method of video compression has been proposed. Neural image compression algorithm is the key component of our method. It is based on a well know method called predictive vector quantization (PVQ). It combines two different techniques: vector quantization and differential pulse code modulation. The neural video compression method based on PVQ algorithm requires correct detection of key frames in order to improve its performance. For key frame detection our method uses techniques based on the Restricted Boltzmann Machine method (RBM).

The research presented in the paper was inspired by the work of R. Douglas Fields. It transpired that not only neural structures in the brain play huge role in the process of understanding but also glial cells, which have so far been treated as passive cells with the task of protecting neuronal cells. This was a motivation to the proposed idea that currently extremely popular convolutional neural networks should be equipped with some elements corresponding to glial cells. In this work we present a modification of convolutional structures, which consist in adding additional adjustable parameters. The parameters control convolutional filter outputs. This approach allowed us to improve the quality of classification. In addition, the newly proposed structure is easier to interpret by indicating which filters are specific to a particular class of visual objects.

Classification of distorted patterns poses real problem for majority of classifiers. In this paper we analyse robustness of deep neural network in classification of such patterns. Using specific convolutional network architecture, an impact of different types of noise on classification accuracy is evaluated. For highly distorted patterns to improve accuracy we propose a preprocessing method of input patterns. Finally, an influence of different types of noise on classification accuracy is also analysed.

This paper is an attempt to design a fast image retrieval system with limited hardware resources. To this end, we use two-stage color-based features, Hadoop with HDFS to ensure file system flexibility, even in the case of sprawling into cloud projects and JAVA environment to run on every operating system. Namely, we retrieve images by color histogram and then by the color coherence vector to pick the best match from the results found by the previous algorithm. We tested the system on a large set of Microsoft COCO images.

The paper presents the intelligent driving assistant system as a device to increase a car active safety without any interference with a driving process. The system - based on the softcomputing methodology working “on-line” - is able to overtake the driver’s reaction. The system analyses pictures in front of the vehicle and recognises road events and the grip of the road. The driver is informed about the each kind of recognised event. To resolve the problem of the road event recognition entirely new picture preprocessing approach has been used. The learning for multilayer perceptron realised by such data gives very good results. The new way of extracting data from pictures is a promising solution. The algorithm was implemented as part of a real system to support the on-line driver decision. The system was tested in the real car in real traffic with very promising results.

This paper proposes a new image descriptor based on color spatial distribution for image similarity comparison. It is similar to methods based on HOG and spatial pyramid but in contrast to them operates on colors and color directions instead of oriented gradients. The presented method assumes using two types of descriptors. The first one is used to describe segments of similar color and the second sub-descriptor describes connections between different adjacent segments. By this means we gain the ability to describe image parts in a more complex way as is in the case of the histogram of oriented gradients (HOG) algorithm but more general as is in the case of keypoint-based methods such as SURF or SIFT. Moreover, in comparison to the keypoint-based methods, the proposed descriptor is less memory demanding and needs only a single step of image data processing. Descriptor comparing is more complicated but allows for descriptor ordering and for avoiding some unnecessary comparison operations.

As an extension of using image segmentation to do stereo matching, firstly, by using self-organizing map (som) and K-means algorithms, this paper provides a self-distributed segmentation method that allocates segments according to image’s texture changement where in most cases depth discontinuities appear. Then, for stereo, under the fact that the segmentation of left image is not exactly same with the segmentation of right image, we provide a matching strategy that matches segments of left image to pixels of right image as well as taking advantage of border information from these segments. Also, to help detect occluded regions, an improved aggregation cost that considers neighbor valid segments and their matching characteristics is provided. For post processing, a gradient border based median filter that considers the closest adjacent valid disparity values instead of all pixels’ disparity values within a rectangle window is provided. As we focus on real-time execution, these time-consumming works for segmentation and stereo matching are executed on a massively parallel cellular matrix GPU computing model. Finaly, we provide our visual dense disparity maps before post processing and final evaluation of sparse results after post-processing to allow comparison with several ranking methods top listed on Middlebury.

A novel Computer Aided Diagnosis System for early diagnosis of Diabetic Retinopathy is proposed for the detection and classification of Bright lesion classes and Dark lesion classes of Fundus Retina images using machine learning mechanisms. In the proposed methodology, the detection procedure is based on Fuzzy C Means (FCM) clustering technique to segment the candidate region areas. In the Dark lesion category, attempts are being made to modify the Micro aneurysms detection and Blood vessel elimination with the help of improvised algorithms. For the classification of each Bright and Dark lesion classes a classification system is built using machine learning algorithms namely Naive Bayes and Support Vector Machine. A comparative study between the two machine learning algorithms yield accuracy of 97.0588 % for Bright lesion classification using Naive Bayes classifier and accuracy of 88.8889 % for Dark lesion classification using Support Vector Machine classifier.

In this paper we present a system intended for content-based image retrieval tightly integrated with a relational database management system. Users can send query images over the appropriate web service channel or construct database queries locally. The presented framework analyses the query image based on descriptors which are generated by the bag-of-features algorithm and local interest points. The system returns the sequence of similar images with a similarity level to the query image. The software was implemented in .NET technology and Microsoft SQL Server 2012. The modular construction allows to customize the system functionality to client needs but it is especially dedicated to business applications. Important advantage of the presented approach is the support by SOA (Service-Oriented Architecture), which allows to use the system in a remote way. It is possible to build software which uses functions of the presented system by communicating over the web service API with the WCF technology.

The paper describes new algorithms proposed for the granular pattern recognition system that retrieves an image from a collection of color digital pictures based on the knowledge contained in the object information granule (OIG). The algorithms use the granulation approach that employs fuzzy and rough granules. The information granules present knowledge concerning attributes of the object to be recognized. Different problems are considered depending on the full or partial knowledge where attributes are “color”, “location”, “size”, “shape”.