Skip to main content

Über dieses Buch

The two-volume set LNAI 10245 and LNAI 10246 constitutes the refereed proceedings of the 16th International Conference on Artificial Intelligence and Soft Computing, ICAISC 2017, held in Zakopane, Poland in June 2017.

The 133 revised full papers presented were carefully reviewed and selected from 274 submissions. The papers included in the first volume are organized in the following five parts: neural networks and their applications; fuzzy systems and their applications; evolutionary algorithms and their applications; computer vision, image and speech analysis; and bioinformatics, biometrics and medical applications.



Neural Networks and Their Applications


Author Profiling with Classification Restricted Boltzmann Machines

This paper discusses author profiling of English-language mails and blogs using Classification Restricted Boltzmann Machines. We propose an author profiling framework with no need for handcrafted features and only minor use of text preprocessing and feature engineering. The classifier achieves competitive results when evaluated with the PAN-AP-13 corpus: 36.59% joint accuracy, 57.83% gender accuracy and 59.17% age accuracy. We also examine the relations between discriminative, generative and hybrid training methods.

Mateusz Antkiewicz, Marcin Kuta, Jacek Kitowski

Parallel Implementation of the Givens Rotations in the Neural Network Learning Algorithm

The paper describes a parallel feed-forward neural network training algorithm based on the QR decomposition with the use of the Givens rotation. The beginning brings a brief mathematical background on Givens rotation matrices and elimination step. Then the error criterion and its necessary transformations for the QR decomposition are presented. The paper’s core holds an essential explanation to accomplish hardware-based parallel implementation. The paper concludes with a theoretical description of speed improvement gained by parallel implementation of the Givens reduction in the QR decomposition process.

Jarosław Bilski, Bartosz Kowalczyk, Jacek M. Żurada

Parallel Levenberg-Marquardt Algorithm Without Error Backpropagation

This paper presents a new parallel architecture of the Leven-berg-Marquardt (LM) algorithm for training fully connected feedforward neural networks, which will also work for MLP but some cells will stay empty. This approach is based on a very interesting idea of learning neural networks without error backpropagation. The presented architecture is based on completely new parallel structures to significantly reduce a very high computational load of the LM algorithm. A full explanation of parallel three-dimensional neural network learning structures is provided.

Jarosław Bilski, Bogdan M. Wilamowski

Spectral Analysis of CNN for Tomato Disease Identification

Although Deep Convolutional Neural Networks have been widely applied for object recognition, most of the works have often based their analysis on the results generated by a specific network without considering how the internal part of the network itself has generated those results. The visualization of the activations and features of the neurons generated by the network can help to determine the best network architecture for our proposed idea. By the application of deconvolutional networks and deep visualization, in this work, we propose an analysis to determine which kind of images with different color spectrum provide better information to generate a better accuracy of our CNN model. The focus of this study is mostly based on the identification of diseases and plagues on plants. Experimental results on images with different diseases from our Tomato disease dataset show that each disease contains valuable information in the infected part of the leaf that responds differently to other uninfected parts of the plant.

Alvaro Fuentes, Dong Hyeok Im, Sook Yoon, Dong Sun Park

From Homogeneous Network to Neural Nets with Fractional Derivative Mechanism

The paper refers to ANNs of the feed-forward type, homogeneous within individual layers. It extends the idea of network modelling and design with the use of calculus of finite differences proposed by Dudek-Dyduch E. and then developed jointly with Tadeusiewicz R. and others. This kind of neural nets was applied mainly to different features extraction i.e. edges, ridges, maxima, extrema and many others that can be defined with the use of classic derivative of any order and their linear combinations. Authors extend this type ANNs modelling by using fractional derivative theory. The formulae for weight distribution functions expressed by means of fractional derivative and its discrete approximation are given. It is also shown that the application of discrete approximation of fractional derivative of some base functions allows for modelling the transfer function of a single neuron for various characteristics. In such an approach smooth control of a derivative order allows to model the neuron dynamics without direct modification of the source code in IT model. The new approach presented in the paper, universalizes the model of the considered type of ANNs.

Zbigniew Gomolka, Ewa Dudek-Dyduch, Yuriy P. Kondratenko

Neurons Can Sort Data Efficiently

This paper introduces an efficient sorting algorithm that uses new models of receptors and neurons which apply the time-conditional approach characteristic for nervous systems. These models have been successfully applied to automatically construct neural graphs that consolidate representation of all sorted objects and relations between them. The introduced parallely working algorithm sorts objects simultaneously for all attributes constructing an active associative neural graph representing all sorted objects in linear time. The sequential version works in linear or sub-linearithmic time. The paper argues that neurons can be used for efficient sorting of objects and the constructed network can be further used to explore relationships between these objects.

Adrian Horzyk

Avoiding Over-Detection: Towards Combined Object Detection and Counting

Existing object detection frameworks in the deep learning field generally over-detect objects, and use non-maximum suppression (NMS) to filter out excess detections, leaving one bounding box per object. This works well so long as the ground-truth bounding boxes do not overlap heavily, as would be the case with objects that partially occlude each other, or are packed densely together. In these cases it would be beneficial, and more elegant, to have a fully end-to-end system that outputs the correct number of objects without requiring a separate NMS stage. In this paper we discuss the challenges involved in solving this problem, and demonstrate preliminary results from a prototype system.

Philip T. G. Jackson, Boguslaw Obara

Echo State Networks Simulation of SIR Distributed Control

Echo State Networks (ESNs) have been shown to be effective for a number of tasks, including motor control, dynamic time series prediction, and memorising musical sequences. In this paper, we propose a new task of ESNs in order to solve distributed optimal control problems for systems governed by parabolic differential equations with discrete time delay using an adaptive critic designs. The optimal control problems are discretised by using a finite element method in time and space, then transcribed into a nonlinear programming problems. To find optimal controls and optimal trajectories ESNs adaptive critic designs are used to approximate co-state equations. The efficiency of our approach is demonstrated for a SIR distributed system to control the spread of diseases.

Tibor Kmet, Maria Kmetova

The Study of Architecture MLP with Linear Neurons in Order to Eliminate the “vanishing Gradient” Problem

Research in deep neural networks are becoming popular in artificial intelligence. Main reason for training difficulties is the problem of vanishing gradients while number of layers increases. While such networks are very powerful they are difficult in training. The paper discusses capabilities of different neural network architectures and presents the proposition of new multilayer architecture with additional linear neurons, that is much easier to train that traditional MLP network and reduces effect of vanishing gradients. Efficiency of suggested approach has been confirmed by several exeriments.

Janusz Kolbusz, Pawel Rozycki, Bogdan M. Wilamowski

Convergence and Rates of Convergence of Recursive Radial Basis Functions Networks in Function Learning and Classification

In this paper we consider convergence and rates of convergence of the normalized recursive radial basis function networks in function learning and classification when network parameters are learned by the empirical risk minimization.

Adam Krzyżak, Marian Partyka

Solar Event Classification Using Deep Convolutional Neural Networks

The recent advances in the field of neural networks, more specifically deep convolutional neural networks (CNN), have considerably improved the performance of computer vision and image recognition systems in domains such as medical imaging, object recognition, and scene characterization. In this work, we present the first attempt into bringing CNNs to the field of Solar Astronomy, with the application of solar event recognition. With the objective of advancing the state-of-the-art in the field, we compare the performance of multiple well established CNN architectures against the current methods of multiple solar event classification. To evaluate the effectiveness of deep learning in the solar image domain, we experimented with well-known architectures such as LeNet-5, CifarNet, AlexNet, and GoogLenet. We investigated the recognition of four solar event types using image regions extracted from the high-resolution full disk images of the Sun from the NASA’s Solar Dynamics Observatory (SDO) mission. This work demonstrates the feasibility of using CNNs by obtaining improved results over the conventional pattern recognition methods used in the field.

Ahmet Kucuk, Juan M. Banda, Rafal A. Angryk

Sequence Learning in Unsupervised and Supervised Vector Quantization Using Hankel Matrices

In the present contribution we consider sequence learning by means of unsupervised and supervised vector quantization, which should be invariant regarding to shifts in the sequences. A mathematical tool to achieve a respective invariant representation and comparison of sequences are Hankel matrices with an appropriate dissimilarity measure based on subspace angles. We discuss their mathematical properties and show how they can be incorporated in prototype based vector quantization schemes like neural gas and self-organizing maps for clustering and data compression in case of unsupervised learning. For classification learning we refer to the closely related supervised learning vector quantization scheme. Particularly, median variants of these vector quantizers allow an easy application of Hankel matrices. A possible application of the Hankel matrix approach could be the analysis of DNA sequences, as it does not require the alignment of sequences due to its invariance properties.

Mohammad Mohammadi, Michael Biehl, Andrea Villmann, Thomas Villmann

Discrete Cosine Transformation as Alternative to Other Methods of Computational Intelligence for Function Approximation

The discrete cosine transform (DCT) is commonly known in signal processing. In this paper DCT is used in computational intelligence to show its usefulness. Proposed DCT method is used to reduce the size of system which results in faster processing with limited and controlled precision lost. Proposed method is compared to other ones like Fuzzy Systems, Neural Networks, Support Vector Machines, etc. to investigate the ability to solve sample problem. The results show that the method can be successfully used and the results are comparable or better to those achieved by other methods considered as powerful ones.

Angelika Olejczak, Janusz Korniak, Bogdan M. Wilamowski

Improvement of RBF Training by Removing of Selected Pattern

Number of training patterns has a huge impact on artificial neural networks training process, not only because of time-consuming aspects but also on network capacities. During training process the error for the most patterns reaches low error very fast and is hold to the end of training so can be safely removed without prejudice to further training process. Skilful removal of patterns during training allow to achieve better training results decreasing both training time and training error. The paper presents some implementations of this approach for Error Correction algorithm and RBF networks. The effectiveness of proposed methods has been confirmed by several experiments.

Pawel Rozycki, Janusz Kolbusz, Oleksandr Lysenko, Bogdan M. Wilamowski

Exploring the Solution Space of the Euclidean Traveling Salesman Problem Using a Kohonen SOM Neural Network

In this paper we present a new approach to solving the Euclidean traveling salesman problem (ETSP) using SOM Kohonen maps with chain topology. The Kohonen learning rule is used with random parameters providing different neuron locations. Any new neuron configuration allows us to obtain a new ETSP solution. This new approach to exploring the solution space of the ETSP is easy to implement and suitable for relatively large ETSP problems. Furthermore, the approach could be combined both with other global optimization methods as genetic algorithms and with simple TSP solving heuristic procedures. The method is illustrated by simulations used for solving some TSPLIB problems.

Ewa Skubalska-Rafajłowicz

Resolution Invariant Neural Classifiers for Dermoscopy Images of Melanoma

This article contributes to the Computer Aided Diagnosis (CAD) of melanoma pigmented skin cancer. We test back-propagated Artificial Neural Network (ANN) classifiers for discrimination in benign and malignant skin lesions. Features used for the classification are derived from wavelet decomposition coefficients of the dermoscopy image. We show the most efficient ANN setups as a function of the structure of hidden layers and the network learning algorithms. Our network topologies are limited for the sake of restrictions in memory and processing power of smartphones which are more and more popular as hand-held ‘mobile’ CAD devices for melanoma. We claim resolution invariance of the selected wavelet features.

Grzegorz Surówka, Maciej Ogorzałek

Application of Stacked Autoencoders to P300 Experimental Data

Deep learning has emerged as a new branch of machine learning in recent years. Some of the related algorithms have been reported to beat state-of-the-art approaches in many applications. The main aim of this paper is to verify one of the deep learning algorithms, specifically a stacked autoencoder, to detect the P300 component. This component, as a specific brain response, is widely used in the systems based on brain-computer interface. A simple brain-computer interface experiment more than 200 school-age participants was performed to obtain large datasets containing the P300 component. After feature extraction the collected data were split into the training and testing sets. State-of-the art BCI classifiers (such as LDA, SVM, or Bayesian LDA) were applied to the data and then compared with the results of stacked autoencoders.

Lukáš Vařeka, Tomáš Prokop, Roman Mouček, Pavel Mautner, Jan Štěbeták

NARX Neural Network for Prediction of Refresh Timeout in PIM–DM Multicast Routing

In this paper, we propose a novel method for optimization of multicast routing. With the use of a NARX neural network, we predict a refresh timeout in PIM–DM algorithm.

Nataliia Vladymyrska, Michał Wróbel, Janusz T. Starczewski, Viktoriia Hnatushenko

Evolving Node Transfer Functions in Deep Neural Networks for Pattern Recognition

Theoretical results suggest that in order to learn complicated functions that can represent high-level features in the computer vision field, one may need to use deep architectures. The popular choice among scientists and engineers for modeling deep architectures are feed-forward Deep Artificial Neural Networks. One of the latest research areas in this field is the evolution of Artificial Neural Networks: NeuroEvolution. This paper explores the effect of evolving a Node Transfer Function and its parameters, along with the evolution of connection weights and an architecture in Deep Neural Networks for Pattern Recognition problems. The results strongly indicate the importance of evolving Node Transfer Functions for shortening the time of training Deep Artificial Neural Networks using NeuroEvolution.

Dmytro Vodianyk, Przemysław Rokita

A Neural Network Circuit Development via Software-Based Learning and Circuit-Based Fine Tuning

A development method of neural network with software-based learning and circuit-based fine tuning is proposed. The backpropagation is known as one of the most efficient learning algorithms. A weakness is that the hardware implementation is extremely difficult. The RWC algorithm which is very easy to implement its hardware circuits takes too many iterations for learning. In the proposed approach, the main learning is performed with a software version of the BP algorithm and then, learned weights are transplanted on a hardware version of a neural circuit. At the time of the weight transplantation, significant amount of output error would occur due to the characteristic difference between the software and the hardware. In the proposed method, such error is reduced via a complementary learning of RWC algorithm which is implemented in a simple hardware.

Changju Yang, Shyam Prasad Adhikari, Michal Strzelecki, Hyongsuk Kim

Fuzzy Systems and Their Applications


A Comparative Study of Two Novel Approaches to the Rule-Base Evidential Reasoning

Recently two new approaches to the rule-base evidential reasoning were proposed in the literature. The first of them is based on the Atannasov’s intuitionistic fuzzy sets theory ($$A-IFS$$) and Dempster-Shafer theory (DST). The second one is based directly on the synthesis of fuzzy logic and DST. In this paper, using the simple, but real-world example, it is shown that the first approach in the critical case of high conflict provides more reasonable and intuitively obvious results.

Ludmila Dymova, Krzysztof Kaczmarek, Pavel Sevastjanov

STRIPS in Some Temporal-Preferential Extension

In 1971 N. Nilson introduced a very smart improvement of forward search methodology in classical planning. It is commonly known as STRIPS. However, the original STRIPS is not sensitive to temporal and preferential aspects of reasoning. Unfortunately, neither temporal, nor preferential extension of STRIPS is known. This paper is just aimed at proposing such an extension, called later TP-STRIPS. In addition, some of its meta-logical properties are proved. It is also shown how TP-STRIPS may be exploited in more practical contexts.

Krystian Jobczyk, Antoni Ligeza

Geometrical Interpretation of Impact of One Set on Another Set

In the paper, we describe new tacit problems during the process of comparison of objects. The direction of objects’ comparison seems to have essential role because such comparison may not be symmetric. Thus, we can say that two objects may be viewed as an attempt to determine the degree to which they are similar or different. In this paper, we consider objects described by a set of nominal attributes which values are not precisely known or can be repeated in the object description. Two kinds of objects’ descriptions are considered, the first the fuzzy and the second the multiset description. Asymmetric phenomena of comparing such descriptions of objects is emphasized and discussed.

Maciej Krawczak, Grażyna Szkatuła

A Method for Nonlinear Fuzzy Modelling Using Population Based Algorithm with Flexibly Selectable Operators

In this paper a new method based on a population-based algorithm with flexible selectable operators for nonlinear modeling is proposed. This method enables usage of any types of exploration and exploitation operators, typical for population-based algorithms. Moreover, in proposed approach each solution from population encodes activity and parameters of these operators. Due to this, they can be selected dynamically in the evolution process. Such approach eliminates the need for determining detailed mechanism of the population-based algorithm. For the simulations typical nonlinear modeling benchmarks were used.

Krystian Łapa, Krzysztof Cpałka, Lipo Wang

Fuzzy Portfolio Diversification with Ordered Fuzzy Numbers

In this paper, we consider a multi-objective portfolio diversification problem under real constraints in fuzzy environment, where the objective is to minimize the variance of portfolio and maximize expected return rate of portfolio. The return rates of assets are modeled using concept of Ordered Fuzzy Candlesticks, which are Ordered Fuzzy Numbers. The use of them allows modeling uncertainty associated with financial data based on high-frequency data. Thanks to well-defined arithmetic of Ordered Fuzzy Numbers, the estimators of fuzzy-valued expected value and covariance can be computed in the same way as for real random variables. In an empirical study, 20 assets included in the Warsaw Stock Exchange Top 20 Index are used to compare considered fuzzy model with crisp mean-variance model.

Adam Marszałek, Tadeusz Burczyński

Using a Hierarchical Fuzzy System for Traffic Lights Control Process

The present study presents the applications of a hierarchical fuzzy system in the traffic light system control process. In this paper a hierarchical fuzzy system was designed. This system is based on the fuzzy system for traffic lights control FS-TLC. The main advantage of the solution presented herein is a significant reduction of the number of defined rules. Additionally, owing to the use of hierarchical fuzzy systems, it is possible to greatly reduce the time required for computations and to reduce the use of memory in the systems designed. Also that systems can be significantly more effective in operation, and they allow a creation of faster controllers that may be implemented in the hardware architecture.

Bartosz Poletajew, Adam Slowik

Hierarchical Fuzzy Logic Systems in Classification: An Application Example

This paper focuses on problems related to learning rules using numerical data for the Hierarchical Fuzzy Logic Systems (HFLS) described in [8]. Using hierarchical structure of Fuzzy Logic Systems (FLS) some complex problems could be divided into subproblems with smaller dimensions. “Hierarchical” means that fuzzy sets produced as output of one of fuzzy logic systems are then processed as input of another as the sets of auxiliary variables. The main scope of this paper is to use HFLS in classification problems for different datasets from the UCI Machine Learning Repository (The UC Irvine Machine Learning Repository shared by Center for Machine Learning and Intelligent Systems (University of California, Irvine) available at The proposal presented in this paper operates on a type-1 HFLS, built with the fuzzy logic systems (in the sense of Mamdani). Iris, Abalone, Wine, Wine Quality Red and White datasets were used. Obtained results are described and compared to other classification systems.

Krzysztof Renkas, Adam Niewiadomski

A Bullying-Severity Identifier Framework Based on Machine Learning and Fuzzy Logic

Bullying at schools is a serious social phenomenon around the world that negatively affects the development of children. However, anti-bullying programs should not focus on labeling children as either bullies or victims since they could produce opposite effects. Thus, an approach to deal with bullying episodes, without labeling children, is to determine their severity, so that school staff may respond to them appropriately. Related work about computational techniques to fight against bullying showed promising results but they offer categorical information as a set of labels. This work proposes a framework to determine bullying severity in texts, composed by two parts: (1) evaluation of texts using Support Vector Machine (SVM) classifiers found in the literature, and (2) development of a Fuzzy Logic System that uses the outputs of SVM classifiers as its inputs to identify the bullying severity. Results show that it is necessary to improve the accuracy of SVM classifiers to determine the bullying severity through Fuzzy Logic.

Carmen R. Sedano, Edson L. Ursini, Paulo S. Martins

Evolutionary Algorithms and Their Applications


On the Efficiency of Successful-Parent Selection Framework in the State-of-the-art Differential Evolution Variants

Successful-parent selection (SPS) framework in differential evolution (DE) is studied. Two SPS versions (SPS1 proposed recently in literature and SPS2 newly proposed in this paper) are applied to seven state-of-the-art DE variants. The algorithms are compared experimentally on CEC 2014 test suite used as a benchmark. The application of SPS1 increases the efficiency of two DE algorithms in over $$50\%$$ of test problems. An overall comparison shows that the newly proposed SPS2 performs well only in two cases whereas SPS1 outperforms six out of 7 original algorithms.

Petr Bujok

State Flipping Based Hyper-Heuristic for Hybridization of Nature Inspired Algorithms

The paper presents a novel hyper-heuristic strategy for hybridization of nature inspired algorithms. The strategy is based on switching the state of agents using a logistic probability function, which depends upon the fitness rank of an agent. A case study using two nature inspired algorithms (Artificial Bee Colony (ABC) and Krill Herding (KH)) and eight optimization problems (Ackley Function, Bukin Function N.6, Griewank Function, Holder Table Function, Levy Function, Schaffer Function N.2, Schwefel Function, Shubert Function) is presented. The results show a superiority of the proposed hyper-heuristic (mean end-rank for hybrid algorithm is 1.435 vs. 2.157 for KH and 2.408 for ABC).

Robertas Damaševičius, Marcin Woźniak

Improved CUDA PSO Based on Global Topology

We introduce a well-optimized implementation of PSO algorithm based on, Compute Unified Device Architecture (CUDA), using global neighborhood topology with extremely large swarms (greater than 1000 particles). The algorithm optimization is based on effective data organization in GPU memory such as transfer and thread optimization, pinned memory and the zero-copy mechanism usage. Experimental results show that the implementation on GPU is significantly faster than implementation on CPU.

Joanna Kołodziejczyk, Dariusz Sychel, Aneta Bera

Optimization of Evolutionary Instance Selection

Evolutionary instance selection is the most accurate process comparing to other methods based on distance, such as the instance selection methods based on k-NN. However, the drawback of evolutionary methods is their very high computational cost. We compare the performance of evolutionary and classical methods and discuss how to minimize the computational cost using optimization of genetic algorithm parameters, joining them with the classical instance selection methods and caching the information used by k-NN.

Mirosław Kordos

Dynamic Difficulty Adjustment for Serious Game Using Modified Evolutionary Algorithm

Dynamic Difficulty Adjustment (DDA) seeks to adapt the challenge a game poses to a human player. When the game is too easy the player can become bored, when it is too hard - frustrated. In the case of a serious game (educational game), additionally, without a balance between the player competence and the game challenge the game could repeatedly exploit the developed skills, or fail to achieve the pedagogical goals. In this paper evolutionary algorithm (EA) is used to find game settings suitable for the player of a serious math game. To reduce the number of training data and accelerate the search for the ‘right’ game difficulty level EA modifications are introduced. Various experiments are performed. The obtained results show that proposed methods can substantially decrease the time a human player has to wait for a suitable game level.

Ewa Lach

Hybrid Initialization in the Process of Evolutionary Learning

Population-based algorithms are an interesting tool for solving optimization problems. Their performance depends not only on their specification but also on methods used for initialization of initial population. In this paper a new hybridization approach of initialization methods is proposed. It is based on classification of initialization methods that allow various combination of the methods from each category. To test the proposed approach typical problems related to population-based algorithms were used.

Krystian Łapa, Krzysztof Cpałka, Yoichi Hayashi

A Tuning of a Fractional Order PID Controller with the Use of Particle Swarm Optimization Method

The paper is devoted to present a new tuning method for Fractional Order PID controller dedicated to temperature control. The proposed method uses Particle Swarm Optimization algorithm. The control plant is described by transfer function with delay. Results of experiments show that the proposed approach assures the good control performance in the sense of known integral cost functions.

Krzysztof Oprzędkiewicz, Klaudia Dziedzic

Controlling Population Size in Differential Evolution by Diversity Mechanism

A new mechanism for resizing population in differential evolution algorithm based on diversity of population has been proposed and compared with linear reduction of population size published in 2014. Seven modifications of differential evolution algorithm were chosen for this comparison. Experiments are done on CEC2014 benchmark set. The new diversity-based resizing mechanism frequently improves results of tested variants of differential evolution algorithm more than linear reduction of population size, especially in larger dimensions.

Radka Poláková

Cosmic Rays Inspired Mutation in Genetic Algorithms

In this paper a new mutation operator is presented. It is based on simulating cosmic ray impact on living tissue. It was proved that the proposed mutation method has a compound probability distribution, which is also derived. Numerical experiments indicate the usefulness of this concept for problems of moderate sizes.

Wojciech Rafajłowicz

OC1-DE: A Differential Evolution Based Approach for Inducing Oblique Decision Trees

This paper describes the application of a Differential Evolution based approach for inducing oblique decision trees in a recursive partitioning strategy. Considering that: (1) the task of finding an optimal hyperplane with real-valued coefficients is a complex optimization problem in a continuous space, and (2) metaheuristics have been successfully applied for solving this type of problems, in this work a differential evolution algorithm is applied with the aim of finding near-optimal hyperplanes that will be used as test conditions of an oblique decision tree. The experimental results show that this approach induces more accurate classifiers than those produced by other proposed induction methods.

Rafael Rivera-Lopez, Juana Canul-Reich, José A. Gámez, José M. Puerta

An Application of Generalized Strength Pareto Evolutionary Algorithm for Finding a Set of Non-Dominated Solutions with High-Spread and Well-Balanced Distribution in the Logistics Facility Location Problem

The paper presents an application of generalized Strength Pareto Evolutionary Algorithm (SPEA) in the Logistic Facilities Location (LFL) problem. The task is to optimize a distribution network, i.e. the number of distribution centers and their locations as well as the number of clients served by the particular centers in terms of three following contrary/contradictory criteria: (a) the total maintenance cost of the network, (b) carbon emissions emitted by combustion engines of trucks into the atmosphere (subjects to minimization) and (c) the customer service reliability (subject to maximization). For this purpose, an original multi-objective optimization technique which allow to obtain a set of so-called non-dominated solutions of the considered problem, representing different levels of compromise between the above criteria, is applied. In order to provide a broad, flexible selection of the final solution from the obtained set, the proposed approach aims at finding the set of solutions with high spread and well-balanced distribution in the objective (criteria) space. The functionality of our technique is demonstrated using numerical experiments. Its distinct advantages over alternative approaches are presented in the frame of comparative analysis as well.

Filip Rudziński

Efficient Creation of Population of Stable Biquad Sections with Predefined Stability Margin for Evolutionary Digital Filter Design Methods

Evolutionary methods are very often used to digital filters design (especially for digital filters with non-standard amplitude characteristics). In practical digital filter implementation, we can use a cascade of biquad sections. Each biquad section will be stable if all poles of transfer function are located into unitary circle in the z-plane. To generate $$k-th$$ stable biquad section, we can use an existing equations which are generated for the case when the coefficient $$a_{k,0}=1$$. The problem is complicated if we want to have the vary values of all filter coefficients from the continuous range $$[-1; 1]$$ or from the discrete range $$[-1; 1-2^{-M}]$$ (if filter will be implemented into Q.M fixed-point format). In this paper we have presented an efficient method for generation of stable biquad sections. The proposed method can be used in any evolutionary digital filter design method for increase its efficiency. Using proposed approach we can fast generate the population of stable biquad sections with prescribed stability margin. Due to presented approach, we can also very fast evaluate the stability of the given biquad section (the methods for polynomial roots generation are not needed).

Adam Slowik

Computer Vision, Image and Speech Analysis


Contiguous Line Segments in the Ulam Spiral: Experiments with Larger Numbers

In our previous papers we have investigated the directional structure and the numbers of straight line segments in the Ulam spiral. Our tests were limited to primes up to $$25\,009\,991$$ due to memory limits. Now we have results for primes up to about $$10^9$$ for the previously used directional resolution, and for the previous maximum number but with greatly increased directional resolution. For the extended resolution, new long segments have been found, among them the first one with 14 points. For larger numbers and the previous resolution, the new segments having up to 13 points were found, but the longest one is still the one with 16 points. It was confirmed that the relation of the number of segments of various lengths to the corresponding number of primes for a given integer, for large numbers, is close to linear in the double logarithmic scale.

Leszek J. Chmielewski, Maciej Janowicz, Grzegorz Gawdzik, Arkadiusz Orłowski

Parallel Realizations of the Iterative Statistical Reconstruction Algorithm for 3D Computed Tomography

The presented paper describes a parallel realization of an approach to the reconstruction problem for 3D spiral x-ray tomography. The reconstruction problem is formulated taking into consideration the statistical properties of signals obtained by x-ray CT and the analytical methodology of image processing. The concept shown here significantly accelerates calculations performed during iterative reconstruction process in the formulated algorithm. Computer simulations have been performed which prove that the reconstruction algorithm described here, does indeed significantly outperform conventional analytical methods in the quality of the images obtained.

Robert Cierniak, Jarosław Bilski, Jacek Smola̧g, Piotr Pluta, Nimit Shah

Efficient Real-Time Background Detection Based on the PCA Subspace Decomposition

We investigate performance of the classical PCA based background subtraction procedure and compare it with the robust PCA versions which are computationally demanding. We show that the simple PCA based version endowed with the fast eigen-decomposition method allows real-time operation on VGA video streams while offering accuracy comparable with some of the robust versions.

Bogusław Cyganek, Michał Woźniak

The Image Classification with Different Types of Image Features

In this paper we present a modified Bag-of-Words algorithm used in image classification. The classic Bag-of-Words algorithm is used in natural language processing. A text (such as a sentence or a document) is represented as a bag of words. In image retrieval or image classification this algorithm also works on one characteristic image feature and most often it is a descriptor defining the surrounding of a keypoint obtained by using e.g. the SURF algorithm. The modification which we have introduced involves using two different types of image features – the descriptor of a keypoint and also the colour histogram, which can be obtained from the surrounding of a keypoint. This additional feature will make it possible to obtain more information as the commonly used SURF algorithm works only on images with greyscale intensity. The experiments which we have conducted show that using this additional image feature significantly improves image classification results by using the BoW algorithm.

Marcin Gabryel, Robertas Damaševičius

Local Keypoint-Based Image Detector with Object Detection

Accurate and efficient image content description is crucial for image retrieval systems. In the paper we propose a novel method to describe images by a combination of the SURF local keypoint detector and the Canny edge detector. Then, a crawler is used to detect objects. The experiments performed on state-of-the-art image dataset showed that the method generates less data than standalone local keypoint detectors.

Rafał Grycuk, Magdalena Scherer, Sviatoslav Voloshynovskiy

Heavy Changes in the Input Flow for Learning Geography of a Robot Environment

A novel approach to generation of geographic knowledge from robot views is presented. It is implemented in a pilot software where a virtual robot operates in a static 2D-environment. The robot sensor scans with rays an angular field of view and produces a 1D view of distances to the closest obstacle. By processing such views, ‘heavy changes’ are detected to trigger switching local maps in an atlas that represents geography of the robot environment. To detect heavy changes, firstly, each plot is transformed to a string of singular points; then, in time-scale, a pair of such strings is subjected to a treatment based on application of the distance of Levenshtein, which leads to so-called Editorial Prescription (EP); a heavy change is detected if EP shows a considerable distinction between strings. This approach is applied in automatic construction of an atlas for non-Cartesian navigation, while robot explores the scene.

Georgii Khachaturov, Josué Figueroa-González, Silvia B. González-Brambila, Juan M. Martínez-Hernández

Open Access

Constant-Time Fourier Moments for Face Detection — Can Accuracy of Haar-Like Features Be Beaten?

We demonstrate a technique allowing for constant-time calculation of low order Fourier moments, applicable in detection tasks. Real and imaginary parts of the moments can be used as features for machine learning and classification of image windows. The technique is based on a set of special integral images, prepared prior to the scanning procedure. The integral images are constructed as cumulative inner products between the input image and suitable trigonometric terms. Additional time invested in the preparation of such integral images is amortized later at the stage of scanning. Then, the extraction of each moment requires only 21 operations, regardless of the number of pixels in the detection window, and thereby is an O(1) calculation.As an application example, face detection experiments are carried out with detectors based on Haar-like features serving as opponents to the proposed Fourier-based detectors.

Przemysław Klȩsk

Neural Video Compression Based on PVQ Algorithm

In this paper we present a video compression algorithm based on predictive vector quantization, which is a combination of vector quantization and differential pulse code modulation. We optimized the algorithm using chroma subsampling which reduces the amount of information that needs to be processed. This allowed us to combine two color channels into one and thereby reduce the number of predictors and codebooks. Furthermore, we introduced inter-frames which only store regions that changed compared to previous frames, further decreasing the size of compressed data.

Michał Knop, Tomasz Kapuściński, Rafał Angryk

Taming the HoG: The Influence of Classifier Choice on Histogram of Oriented Gradients Person Detector Performance

Histogram of oriented gradients (HoG) is a common choice for hand-crafted feature used in a wide range of machine vision task. It functions as a part of a processing pipeline, in which it’s followed by a classifier. The canonical approach proposed by the authors of HoG is the use of a linear support vector machine (SVM). This approach is usually followed by the majority of adopters with good results. However, a range of classifiers have proven to perform better than linear SVM in a variety of applications. In this paper, we investigate the pairing between HoG and a range of classifiers in order to find one with the best performance in terms of accuracy and processing speed for the task of human silhouete detection.

Michał Olejniczak, Marek Kraft

Virtual Cameras and Stereoscopic Imaging for the Supervision of Industrial Processes

In this article we present a concept of visualisation of some industrial processes, such as glass melting. The process is observed from a single camera at fixed location but when the scene geometry is known, like for example in case of glass furnaces, the image can be transformed to a perspective view from a virtual camera at arbitrary location and to stereoscopic views based on two virtual cameras. We applied automatic analysis of the input image to present to the supervisor of the process a synthetic image, which only contains features that are important for the process control. We have developed a prototype visualization system, tested in cooperation with the glass industry.

Paweł Rotter

Object Detection with Few Training Data: Detection of Subsiding Troughs in SAR Interferograms

Subsiding troughs that are the result of mining activities can be detected in SAR interferograms as approximately elliptic shapes against the noisy background. Despite large areas being covered by interferogram, the number of positive samples, which can be used for automatic learning, is limited. In this paper we propose two alternative methods for the detection of subsiding troughs: the first one is designed to detect any circular shapes and does not require any learning set and the second is based on automatic learning but requires a reduced number of positive samples. The two proposed methods can support manual inspection of large areas in SAR interferograms.

Paweł Rotter, Jacek Strzelczyk, Stanisława Porzycka-Strzelczyk, Claudio Feijoo

FPGA-Based System for Fast Image Segmentation Inspired by the Network of Synchronized Oscillators

This paper presents an FPGA-based system for fast and parallel image segmentation. Implemented segmentation method is inspired by operation of the network of synchronized oscillators - a robust tool for image processing and analysis. The architecture of parallel digital image processor was presented and discussed. It was optimized to enable fully synchronized parallel processing along with reduction of FPGA resources. The developed system is able to analyze both binary and monochrome images with size of 64$$\,\times \,$$64 pixels. It was demonstrated that it can perform region growing image segmentation, edge detection, labelling of binary objects, and basic morphological operations. Sample analysis results were also presented and discussed.

Michal Strzelecki, Przemyslaw Brylski, H. Kim

From Pattern Recognition to Image Understanding

This paper shows the trend in the transformation of the classic image recognition via the interpretation of the image content towards automatic shape and image understanding. The approach presented combines the mechanism proposed by Tadeusiewicz in [1] with the theory of granular computing introduced by Pedrycz in [2]. Its name, active partitions, is related to active contour techniques, from which it originates. It provides the ability to transfer the well-known concepts of object localization from the pixel level to image representations with meaningful image granules. Thus, the approach offers a great potential for the development of human-like image content interpretation.

Piotr S. Szczepaniak, Arkadiusz Tomczyk

Linguistic Description of Color Images Generated by a Granular Recognition System

The paper proposes a new method employed in an intelligent pattern recognition system that generates linguistic description of color digital images. The linguistic description is produced based on fuzzy rules and information granules concerning colors as most important among image attributes. With regard to the color, the CIE chromaticity color model is applied, with the concept of fuzzy color areas. The linguistic description uses information about location of color granules in input images.

Krzysztof Wiaderek, Danuta Rutkowska, Elisabeth Rakus-Andersson

Bioinformatics, Biometrics and Medical Applications


Classification of Physiological Data for Emotion Recognition

Emotion recognition is seen to be important not only for computer science or sport activity but also for old and sick people to live independently in their own homes as long as possible. In this paper Empatica E4 wristband is used to collect the date and assess the stress level of the user. We describe an algorithm for the classification of physiological data for emotion recognition. The algorithm has been divided into the following steps: data acquisition, signal preprocessing, feature extraction, and classification. The data acquired during various daily activities consist of more than 3 h of wristband signal. Through various stress tests we achieve a maximum accuracy of 71% for a stressed/relaxed classification. These results lead to the conclusion that Empatica E4 wristband can be used as a device for emotion recognition.

Philip Gouverneur, Joanna Jaworek-Korjakowska, Lukas Köping, Kimiaki Shirahama, Pawel Kleczek, Marcin Grzegorzek

Biomimetic Decision Making in a Multisensor Assisted Living Environment

Various sensors were found adequate to monitor human behavior in natural habitat. Besides metrological factors they were adopted to specific conditions of unobtrusive acquisition in human (e.g. an elder or child). An initial approach focused on use of a single sensor to detect a particular event evolved to behavioral studies based on complex recordings in multisensor environments. In such environment sensors based on different physical principles play complementary role and the resultant detection outperforms any of single component-based if combines the detail information correctly. We follow the rules of neural modulation in human sensory system to propose a biomimetic decision making in multisensor assisted living environment. In our approach each sensor contributes to the final detection accordingly to its presumed reliability in particular human behavior. Moreover, the system learns human habits, predicts most probable future actions from the history and anticipates accordingly the importance of particular sensors.

Piotr Augustyniak, Magdalenia Smoleń

Classification of Splice-Junction DNA Sequences Using Multi-objective Genetic-Fuzzy Optimization Techniques

The main goal of this paper is the application of our fuzzy rule-based classification technique with genetically optimized accuracy-interpretability trade-off to the classification of the splice-junction DNA sequences coming from the Molecular Biology (Splice-junction Gene Sequences) benchmark data set (available from the UCI repository). Two multi-objective evolutionary optimization algorithms are employed and compared in the framework of our technique, i.e., the well-known Strength Pareto Evolutionary Algorithm 2 (SPEA2) and our SPEA2’s generalization (referred to as SPEA3) characterized by a higher spread and a better-balanced distribution of solutions. A comparative analysis with 15 alternative approaches is also performed.

Marian B. Gorzałczany, Filip Rudziński

Automatic Detection of Blue-Whitish Veil as the Primary Dermoscopic Feature

Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Owing to the difficulty and subjectivity of human interpretation, dermoscopy image analysis has become an important research area. One of the most important local structure that is likely to appear in malignant melanoma is the blue-whitish veil. In this article, we present an unsupervised approach to the blue-whitish veil detection in dermoscopy images of pigmented skin lesions based on the analysis of HSV color space. The method is tested on a set of 179 dermoscopy images and the detection error rate is lower than 15%. The results demonstrate that the presented method achieves both fast and accurate blue structure segmentation in dermoscopy images.

Joanna Jaworek-Korjakowska, Paweł Kłeczek, Marcin Grzegorzek, Kimiaki Shirahama

Bio-inspired Topology of Wearable Sensor Fusion for Telemedical Application

Application of wearable sensors is a promising approach in building novel telemedical services. In this paper, we propose the biologically inspired method for monitoring human activity in living conditions. The solution is based on the set of sensors integrated in the single wearable device and imitates the natural arrangement of human perception system. The designed wearable device enables to acquire physiological and environmental parameters. With the use of proposed appliance it is possible to collect body and ambient temperature, barometric pressure, light intensity and acceleration. In the experimental part, the signals were recorded during selected activities of daily living (ADL). The sitting activity classification was implemented using perceptron.

Eliasz Kantoch, Dominik Grochala, Marcin Kajor

An Evaluation of Fuzzy Measure for Face Recognition

In this paper, we analyze the properties and performance of the Choquet integral and fuzzy measure, particularly $$\lambda $$–fuzzy measure in the context of an aggregation of classifiers based on various facial areas. The fuzzy measure and Choquet integral have been shown to be an efficient aggregation techniques. However, in practice reported so far, the choice of the initial values of the measure corresponding to the saliency of facial features has been dependent upon the expert decision. Here, we propose an algorithmic way of finding these values. For this purpose a Particle Swarm Optimization (PSO) method is considered. The reported experimental results show that the method is more effective than the expert – centered approach.

Paweł Karczmarek, Adam Kiersztyn, Witold Pedrycz

Analysis of Dermatoses Using Segmentation and Color Hue in Reference to Skin Lesions

In dermatology there are several well known algorithms of melanocytic lesions recognition but there are not automated algorithms of skin lesion identification and classification. The main aim of this paper is to examine skin changes based on the skin analysis in the chosen model color spaces. With the help of that analysis, the authors show how to extract information from the skin images that will be useful for a future dermatology expert system. In the paper, the authors introduce a novel clinical feature extraction and segmentation method based on modified dermatologists’ approach to diagnose skin lesions. We have also prepared a database (DermDB) of dermoscopic images with the reference data prepared and validated by expert dermatologists.

Lukasz Was, Piotr Milczarski, Zofia Stawska, Marcin Wyczechowski, Marek Kot, Slawomir Wiak, Anna Wozniacka, Lukasz Pietrzak

Improving Data Locality of RNA Secondary Structure Prediction Code

An approach allowing us to improve the locality of RNA Secondary Structure Prediction code is proposed. We discuss the application of this technique to automatic loop nest tiling for the Nussinov algorithm. The approach requires an exact representation of dependences in the form of tuple relations and calculating the transitive closure of dependence graphs. First, to improve code locality, 3-d rectangular tiles are formed within the 3-d iteration space of the Nussinov loop nest. Then tiles are corrected to establish code validity by means of applying the exact transitive closure of a dependence graph. The approach has been implemented as a part of the polyhedral TRACO compiler. The experimental results presents the speed-up factor of optimized code. Related work and future tasks are outlined.

Marek Palkowski, Wlodzimierz Bielecki, Piotr Skotnicki

Robust Detection of Systolic Peaks in Arterial Blood Pressure Signal

The heart rate signal is one of the most important physiological signals characterizing the human heart. The heart bits are usually determined on the basis of the electrocardiographic (ECG) signal. However, they can be also detected by monitoring systolic peaks in a arterial blood pressure (ABP) signal. The pressure signal, as other physiological signals, may be disturbed with noise. In this work we propose the method of precise location of the systolic peaks in ABP signal in the presence of noise, by applying the detection function waveform and fuzzy clustering. The new method is tested using real signals from the MIT-BIH Polysomnographic Database. The results obtained during experiments show the high effectiveness of the proposed method in relation to reference methods.

Tomasz Pander, Robert Czabański, Tomasz Przybyła, Stanisław Pietraszek, Michał Jeżewski

Fuzzy System as an Assessment Tool for Analysis of the Health-Related Quality of Life for the People After Stroke

Stroke remains one of the leading causes of long-term disability in both developed and developing countries. Prevalence and impact of the stroke-related disability on Health-Related Quality of Life (HRQoL) as a recognized and important outcome after stroke is huge. Quick, valid and reliable assessment of the HRQoL in people after stroke constitutes a significant worldwide problem for scientists and clinicians - there are many tools, but no one fulfills all requirements or has prevailing advantages. This paper presents proposition of an evaluation of HRQoL based on the two-level hierarchical fuzzy system. It uses five clinical scores and scales as the inputs and gives in result value from the interval [0; 1]. It may constitute a useful semi-automated tool for supplementary initial assessment of patient functioning and further cyclic re-assessment for rehabilitation process and patient-centered goals of rehabilitation shaping purposes.

Piotr Prokopowicz, Dariusz Mikołajewski, Emilia Mikołajewska, Piotr Kotlarz

Exploratory Analysis of Quality Assessment of Putative Intrinsic Disorder in Proteins

Intrinsically disorder proteins are abundant in nature and can be accurately identified from sequences using computational predictors. While predictions of disorder are relatively easy to obtain there are no tools to assess their quality for a particular amino acid or protein. Quality assessment (QA) scores that quantify correctness of the predictions are not available. We define QA for the prediction of intrinsic disorder and use a large dataset of over 25 thousand proteins and ten modern predictors of disorder to empirically assess the first approach to quantify QA scores. We formulate the QA scores based on the readily available propensities of the intrinsic disorder generated by the ten methods. Our evaluation reveals that these QA scores offer good predictive performance for native structured residues (AUC > 0.74) and poor predictive performance for native disordered residues (AUC < 0.67). Specifically, we show that most of the native disordered residues that are incorrectly predicted as structured have high QA values that inaccurately suggest that these predictions are correct. Consequently, more research is needed to develop high-quality QA scores. We also outline three possible future research directions.

Zhonghua Wu, Gang Hu, Kui Wang, Lukasz Kurgan

Stability Evaluation of the Dynamic Signature Partitions Over Time

Analysis of biometric attributes’ changes is an important issue of behavioral biometrics. It seems to be very important in the case of identity verification. In this paper the analysis of features describing the dynamic signature was performed. The dynamic signature is represented by a set of nonlinear waveforms describing dynamics of signing process. The proposed analysis is based on a set of coefficients defined in the context of the dynamic signature partitioning. The partitioning is performed in order to facilitate analysis of the signature. It consists in division of the signature into parts which can be related to e.g. high and low velocity of pen in the initial and final phase of signing. The proposed method was tested using ATVS-SLT DB dynamic signature database.

Marcin Zalasiński, Krzysztof Cpałka, Meng Joo Er

A Method for Genetic Selection of the Most Characteristic Descriptors of the Dynamic Signature

Dynamic signature verification is an important area of biometrics. In this area methods from the field of computational intelligence can be used. In this paper we propose a new method for genetic selection of the most characteristic descriptors of the dynamic signature. The descriptors are global features of the signature and components created within its partitions. Selection of the descriptors is realized individually for each user of the biometric system. Its purpose is to increase the precision of the biometric system by eliminating the descriptors which do not increase efficiency of verification procedure. Number of descriptors (their combination) can be high, so the use of genetic algorithm to reduce their number seems to be justified. Moreover, reduction of descriptors increases interpretability of fuzzy mechanism for evaluation of signatures’ similarity. Proposed method was tested using known dynamic signatures database-MCYT-100.

Marcin Zalasiński, Krzysztof Cpałka, Yoichi Hayashi

A Method for Changes Prediction of the Dynamic Signature Global Features over Time

Dynamic signature can be represented by a set of global features. These features are interpreted as e.g. number of pen ups, time of signing process, etc. Values of global features can be determined on the basis of non-linear waveforms defining dynamics of the signature. They are acquired using graphic tablet or a device with a touch screen. In this paper we present a method for prediction values of the dynamic signature global features changing over time. The purpose of the prediction is, among others, improving the efficiency of the dynamic signature verification process when the interval between acquisition sessions is large. This interval causes a slight change in the way of signing, which can affect change in the value of global features. In this case the effectiveness of the signature verification also changes (decreases). The possibility of predicting the values of global features can result in a partial elimination of the described problem. Tests of the proposed method were performed using ATVS-SLT DB database of the dynamic signatures.

Marcin Zalasiński, Krystian Łapa, Krzysztof Cpałka, Takamichi Saito


Weitere Informationen