Proceedings of the 9th International Conference on Computer Recognition Systems CORES 2015

The task of classification with imbalanced datasets have attracted quite interest from researchers in the last years. The reason behind this fact is that many applications and real problems present this feature, causing standard learning algorithms not reaching the expected performance. Accordingly, many approaches have been designed to address this problem from different perspectives, i.e., data preprocessing, algorithmic modification, and cost-sensitive learning. The extension of the former techniques to ensembles of classifiers has shown to be very effective in terms of quality of the output models. However, the optimal value for the number of classifiers in the pool cannot be known a priori, which can alter the behaviour of the system. For this reason, ordering-based pruning techniques have been proposed to address this issue in standard classifier learning problems. The hitch is that those metrics are not designed specifically for imbalanced classification, thus hindering the performance in this context. In this work, we propose two novel adaptations for ordering-based pruning metrics in imbalanced classification, specifically the margin distance minimization and the boosting-based approach. Throughout a complete experimental study, our analysis shows the goodness of both schemes in contrast with the unpruned ensembles and the standard pruning metrics in Bagging-based ensembles.

The core idea of the proposed algorithm is to embed the considered dataset into a metric space. Two spaces for embedding of nominal part with the Hamming metric are considered: Euclidean space (the classical approach) and the standard unit sphere $$\mathbb S$$S (our new approach). We proved that the distortion of embedding into the unit sphere is at least 75 % better than that of the classical approach. In our model, combinations of continuous and nominal data are interpreted as points of a cylinder $$\mathbb R^p\times \mathbb S$$Rp×S, where p is the dimension of continuous data. We use a version of the gradient algorithm to compute centroids of finite sets on a cylinder. Experimental results show certain advances of the new algorithm. Specifically, it produces better clusters in tests with predefined groups.

Learning from imbalanced data poses significant challenges for machine learning algorithms, as they need to deal with uneven distribution of examples in the training set. As standard classifiers will be biased toward the majority class there exist a need for specific methods than can overcome this single-class dominance. Most of works concentrated on binary problems, where majority and minority class can be distinguished. But a more challenging problem arises when imbalance is present within multi-class datasets, as relations between classes tend to complicate. One class can be a minority class for some, while a majority for others. In this paper, we propose an efficient method for handling such scenarios that combines the problem decomposition with ensemble learning. According to divide-and-conquer rule, we decompose our multi-class data into a number of binary subproblems using one-versus-one approach. To each simplified task we delegate a ensemble of classifiers dedicated to binary imbalanced problems. Then using a dedicated classifier fusion approach, we reconstruct the original multi-class problem. Experimental analysis backed-up with statistical testing clearly proves that such an approach is superior to state-of-the art ad hoc and decomposition methods used in the literature.

Support Vector Machine (SVM) is a binary classifier, but most of the problems we find in the real-life applications are multiclass. There are many methods of decomposition such a task into the set of smaller classification problems involving two classes only. Two of the widely known are one-versus-one and one-versus-rest strategies. There are several papers dealing with these methods, improving and comparing them. In this paper, we try to combine theses strategies to exploit their strong aspects to achieve better performance. As the performance we understand both recognition ratio and the speed of the proposed algorithm. We used SVM classifier on several different databases to test our solution. The results show that we obtain better recognition ratio on all tested databases. Moreover, the proposed method turns out to be much more efficient than the original one-versus-one strategy.

The main objective of discretization is to transform numerical attributes into discrete ones. The intention is to provide the possibility to use some learning algorithms which require discrete data as input and to help the experts to understand the data more easily. Due to the fact that in classification problems there are high interactions among multiple attributes, we propose the use of evolutionary algorithms to select a subset of cut points for multivariate discretization based on a wrapper fitness function. The algorithm proposed has been compared with the best state-of-the-art discretizers with two decision trees-based classifiers: C4.5 and PUBLIC. The results reported indicate that our proposal outperforms the rest of the discretizers in terms of accuracy and requiring a lower number of intervals.

Extreme learning machine is a new scheme for learning the feedforward neural network, where the input weights and biases determining the nonlinear feature mapping are initiated randomly and are not learned. In this work, we analyze approximation ability of the extreme learning machine depending on the activation function type and ranges from which input weights and biases are randomly generated. The studies are performed on the example of approximation of one variable function with varying complexity. The ranges of input weights and biases are determined for ensuring the sufficient flexibility of the set of activation functions to approximate the target function in the input interval.

Neutrinos are one of the least known elementary particles. The detection of neutrinos is an extremely difficult task since they are affected only by weak subatomic force or gravity. Therefore, large detectors are constructed to reveal neutrino’s properties. Among them the Liquid Argon Time Projection Chamber (LAr-TPC) detectors provide excellent imaging and particle identification ability for studying neutrinos. The computerized methods for automatic reconstruction and identification of particles are needed to fully exploit the potential of the LAr-TPC technique. Herein, the novel method for electron neutrino classification is presented. The method constructs a feature descriptor from images of observed event. It characterizes the signal distribution propagated from vertex of interest, where the particle interacts with the detector medium. The classifier is learned with a constructed feature descriptor to decide whether the images represent the electron neutrino or cascade produced by photons. The proposed approach assumes that the position of primary interaction vertex is known. The method’s performance in dependency to the noise in a primary vertex position and deposited energy of particles is studied.

The main objective of this paper is a texture-based solution to the problem of acute stroke tissue recognition on computed tomography images. Our proposed method of early stroke indication was based on two fundamental steps: (i) segmentation of potential areas with distorted brain tissue (selection of regions of interest), and (ii) acute stroke tissue recognition by extracting and then classifying a set of well-differentiating features. The proposed solution used various numerical image descriptors determined in several image transformation domains: 2D Fourier domain, polar 2D Fourier domain, and multiscale domains (i.e., wavelet, complex wavelet, and contourlet domain). The obtained results indicate the possibility of relatively effective detection of early stroke symptoms in CT images. Selected normal or pathological blocks were classified by LogitBoost with the accuracy close to 75 % with the use of our adjusted cross-validation procedure.

This paper showsMroczek, T.Hippe, Z.S. a new method of explaining Bayesian networks by creating descriptions of their properties in a manner closer to the human perceptual abilities, i.e., decision rules in the IF...THEN form (called by us belief rules). The conversion method is based on the cause and effect analysis of the Bayesian network quantitative component (the probability distribution). Proposed analysis of the quantitative component leads to a deeper insight into the structure of knowledge hidden in the analyzed data set.

Classification of gene expression data is the common denominator of various biomedical recognition tasks. However, obtaining class labels for large training samples may be difficult or even impossible in many cases. Therefore, semi-supervised classification techniques are required as semi-supervised classifiers take advantage of the unlabeled data. Furthermore, gene expression data is high dimensional which gives rise to the phenomena known under the umbrella of the curse of dimensionality, one of its recently explored aspects being the presence of hubs or hubness for short. Therefore, hubness-aware classifiers were developed recently, such as Naive Hubness Bayesian k-Nearest Neighbor (NHBNN). In this paper, we propose a semi-supervised extension of NHBNN and show in experiments on publicly available gene expression data that the proposed classifier outperforms all its examined competitors.

Is it true that everybody knows how to compare classifiers in terms of reliability? Probably not, since it is so common that just after reading a paper we feel that the classifiers’ performance analysis is not exhaustive and we would like to see more information or more trustworthy information. The goal of this paper is to propose a method of multi-classifier comparison on several benchmark data sets. The proposed method is trustworthy, deeper, and more informative (multi-aspect). Thanks to this method, we can see much more than overall performance. Today, we need methods which not only answer the question whether a given method is the best, because it almost never is. Apart from the general strength assessment of a learning machine we need to know when (and whether) its performance is outstanding or whether its performance is unique.

This paper addresses the problem of initialization of the expectation-maximization (EM) algorithm for maximum likelihood estimation of Gaussian mixture models. In order to avoid local maxima of the likelihood function, a genetic algorithm (GA) which searches for best initial conditions of the EM algorithm is proposed. In the GA, a chromosome represents a set of initial conditions, in which initial mean vectors of mixture components are feature vectors chosen from the training set. The chromosome also encodes variances of initial spherical covariance matrices of mixture components. To evaluate each chromosome in the GA we run the EM algorithm until convergence and use the obtained log likelihood as the fitness. In computational experiments our approach was applied to clustering problem and tested on two datasets from the image processing domain. The results indicate that our method outperforms the standard multiple restart EM algorithm and is at least comparable to the state-of-the art random swap EM method.

Fuzzy rule-based systems constitute a wide spread tool for classification problems, but several proposals may decrease its performance when dealing with multi-class problems. Among existing approaches, the FARC-HD algorithm has excelled as it has shown to achieve accurate and compact classifiers, even in the context of multi-class problems. In this work, we aim to go one step further to improve the behavior of the former algorithm by means of a “divide-and-conquer” approach, via binarization in a one-versus-one scheme. Besides, we will contextualize each binary classifier by adapting the database for each subproblem by means of a granularity learning process to adapt the number of fuzzy labels per variable. Our experimental study, using several datasets from KEEL dataset repository, shows the goodness of the proposed methodology.

In this paper we propose a new methodology for evaluating prediction intervals (PIs). TypicallyAlmeida, V., PIs areGama, J. evaluated with reference to confidence values. However, other metrics should be considered, since high values are associated to too wide intervals that convey little information and are of no use for decision-making. We propose to compare the error distribution (predictions out of the interval) and the maximum mean absolute error (MAE) allowed by the confidence limits. Along this paper PIs based on neural networks for short-term load forecast are compared using two different strategies: (1) dual perturb and combine (DPC) algorithm and (2) conformal prediction. We demonstrated that depending on the real scenario (e.g., time of day) different algorithms perform better. The main contribution is the identification of high uncertainty levels in forecast that can guide the decision-makers to avoid the selection of risky actions under uncertain conditions. Small errors mean that decisions can be made more confidently with less chance of confronting a future unexpected condition.

In this paper, a method of elongated structure detection is presented. In general, this is not a trivial task since standard image segmentation techniques require usually quite complex procedures to incorporate the information about the expected shape of the segments. The presented approach may be an interesting alternative for them. In its first phase, it changes the representation of the image. Instead of a set of pixels, the image is described by a set of ellipses representing fragments of the regions of similar color. This representation is obtained using cross-entropy clustering (CEC) method. The second phase analyses geometrical and spatial relationships between ellipses to select those of them that form an elongated structure within an acceptable range of its width. Both phases are elements of hierarchical active partition framework which iteratively collects semantic information about image content.

In the proposed paper, we consider a problem of text detection in document images. This problem plays an important role in OCR systems and is a challenging task. In the first step of our proposed text detection approach, we use a self-adjusting bottom-up segmentation algorithm to segment a document image into a set of connected components (CCs). The segmentation algorithm is based on the Sobel edge detection method. In the second step, CCs are described in terms of 27 features and a machine learning algorithm is then used to classify the CCs as text or nontext. For testing the approach, we have collected a dataset (ASTRoID), which contains 500 images of text blocks and 500 images of nontext blocks. We empirically compare performance of the proposed text detection method when using seven different machine learning algorithms.

This paper presents the improved method for 24 h ahead load forecasting applied to individual household data from a smart metering system. In this approach we decompose a set of individual forecasts into basis latent components with destructive or constructive impact on the prediction. The main research problem in such model aggregation is the proper identification of destructive components that can be treated as some noise factors. To assess the randomness of signals and thus their similarity to the noise, we used a new variability measure that helps to compare decomposed signals with some typical noise models. The experiments performed on individual household electricity consumption data with blind separation algorithms contributed to forecasts improvements.

A hierarchical clustering algorithm based on Gaussian mixture model is presented. The key difference to regular hierarchical mixture models is the ability to store objects in both terminal and nonterminal nodes. Upper levels of the hierarchy contain sparsely distributed objects, while lower levels contain densely represented ones. As it was shown by experiments, this ability helps in noise detection (modeling). Furthermore, compared to regular hierarchical mixture model, the presented method generates more compact dendrograms with higher quality measured by adopted F-measure.

A new model of learning classifier system is introduced to explore continuous-valued environment. The approach applies the real-valued anticipatory classifier system (rACS). In order to handle real-valued inputs effectively, the ternary representation has been replaced by an approach where the interval of real numbers is represented by a natural number. The rACS model has been tested on the 1D linear corridor and the 2D continuous gridworld environments. We show that modified ACS can evolve compact populations of classifiers which represent the optimal solution to the continuous problem.

Computer-assisted detection and segmentation of blood vessels in retinal images of pathological subjects is difficult problem due to the great variability of the images. In this paper we propose an interactive image segmentation system using active learning which will allow quick volume segmentation requiring minimal intervention of the human operator. The advantage of this approach is that it can cope with large variability in images with minimal effort. The collection of image features used for this approach is simple statistics and undirected morphological operators computed on the green component of the image. Image segmentation is produced by classification by a random forest (RF) classifier. An initial RF classifier is built from seed set of labeled points. The human operator is presented with the most uncertain unlabeled voxels to select some of them for inclusion in the training set, retraining the RF classifier. We apply this approach to a well-known benchmarking dataset achieving results comparable to the state of the art in the literature.

The column subset selection problem is a well-known complex optimization problem that has a number of appealing real-world applications including network and data sampling, dimension reduction, and feature selection. There are a number of traditional deterministic and randomized heuristic algorithms for this problem. Recently, it has been tackled by a variety of bio-inspired and evolutionary methods. In this work, differential evolution, a popular and successful real-parameter optimization algorithm, adapted for fixed-length subset selection, is used to find solutions to the column subset selection problem. Its results are compared to a recent genetic algorithm designed for the same purpose.

Data classification is one of the basic tasks in data mining. In this paper, we propose a new classifier based on relative entropy, where data to particular class assignment is made by the majority good guess criteria. The presented approach is intended to be used when relations between datasets and assignment classes are rather complex, nonlinear, or with logical inconsistencies; because such datasets can be too complex to be classified by ordinary methods of decision trees or by the tools of logical analysis. The relative entropy evaluation of associative rules can be simple to interpret and offers better comprehensibility in comparison to decision trees and artificial neural networks.

In the paper a concept of object recognition based on their similarity assessment in case of nonhomogenous qualitative and quantitative objects’ features is presented. Moreover, it is assumed that the features’ intensity values are not given directly but by their pairwise comparative assessment. This corresponds to an intuitive, on human experience-based assessment of the objects’ properties. The proposed object recognition method is based on reference sets divided into credibility layers, according to a relative logical model and conceptual classes of similarity. This concept is illustrated by an example of a conceptual class of “irregular” objects, the “irregularity” being intuitively assessed. The method is presented in the form of an algorithm.

Data clustering is a technique of finding similar characteristics among the data sets which are always hidden in nature, and dividing them into groups. The major factor influencing cluster validation is choosing the optimal number of clusters. A novel random algorithm for estimating the optimal number of clusters is introduced here. The efficiency hybrid random algorithm for good k and modified classical k-means data clustering method in cotton textile imports country clustering and ranking is described and implemented on real-world data set. The original real-world U.S. cotton textile and apparel imports data set is taken under view in this research.

The ensemble selection is one of the important problems in building multiple classifier systems (MCSs). This paper presents dynamic ensemble selection based on the analysis of discriminant functions. The idea of the selection is presented on the basis of binary classification tasks. The paper presents two approaches: one takes into account the normalization of the discrimination functions, and in the second approach, normalization is not performed. The reported results based on the data sets form the UCI repository show that the proposed ensemble selection is a promising method for the development of MCSs.

This paper deals with the comparison of two different approaches for multi-task pattern recognition problem—multi-label and multi-perspective. The experiment performed measured the hamming loss and mean accuracy of both classifiers, to judge which of these two better fit to this kind of problem.

The article concerns the problem of imbalanced data classification. The algorithm improving a standard SMOTE method has been proposed and tested. It is a synergy of the existing approaches and was designed to be more versatile than other similar solutions. To measure the distance between objects, the Euclidean or the HVDM metrics were applied, depending on the number of nominal attributes in a data set.

The goal of this paper is to present a syllable repetition detection method based on linear prediction coefficients obtained by the Levinson–Durbin method. The algorithm wrought by the authors of this paper is based on the linear prediction spectrum. At first the utterance is automatically split into continuous fragments that correspond with syllables. Next, for each of them the formant maps are being obtained. After dimension reduction by the K-means method they are being compared. The algorithm was verified based on 56 continuous utterances of 14 stutterers. They contain fluent parts, as well as syllable repetitions on Polish phonemes. The classifying success reached 90 % of sensitivity with 75–80% precision.

Local descriptors have been one of the most intensively examined mechanisms of image analysis. In this paper, we propose a new chain code-based local descriptor. Unlike many other descriptors existing in the literature, this descriptor is based on string values, which are obtained when starting from a particular point of the image and searching for extrema in a given neighborhood and memorizing a path being traversed through the consequent pixels of the image. We demonstrate that this approach is efficient and helps us preserve both local and global properties of the object. To compare the words we apply the Levenshtein distance. Moreover, four similarity measures (correlation, histogram intersection, chi-square, and Hellinger) are used to compare the histograms of words in the process of classification.

This paper presents an automatic face recognition system, which bases on two-dimensional hidden Markov models. The traditional HMM uses one-dimensional data vectors, which is a drawback in the case of 2D image processing, as part of the information is lost during conversion. The paper presents the full ergodic 2D-HMM and uses it to identify faces. The experimental results demonstrate that the system, basing on two-dimensional hidden Markov models, is able to achieving an average recognition rate of 94 %.

Methods for automatedBoukhers, Z. eye blinkingJarzyński, T. analysis canSchmidt, F. be appliedTiebe, O. to supportGrzegorzek, M. people with certain disabilities in interaction with technical systems, to analyse human deceptive behaviour, in driver fatigue assessment, etc. In this paper we introduce a robust shape-based algorithm for automatic eye blinking detection in video sequences. First, all video frames are classified separately into those showing an open and those corresponding to a closed eye. Second, these classification results are cleverly combined for blinking detection so that the influence of single misclassified frames gets compensated almost completely. In addition to that, we present our investigations on the user behaviour in terms of eye blinking frequency in two different everyday life situations. The most relevant scientific contributions of this paper are (1) the introduction of a new and robust feature extraction technique for the representation of images displaying eyes, (2) a smart fusion scheme improving the results for single-frame classification and (3) the compensation of wrong classification results for single frames providing an almost perfect eye blinking detection rate.

Lip print examination is a very difficult and complex task even for modern forensic departments. Computer systems that will assist a crime scene expert in identification of this kind of evidence are very desired in the forensic science community. Unfortunately, such softwares do not exist as methods of automatic lip print identification are still insufficiently developed. This paper presents an original method of lip print pattern extraction that can be used as a preprocessing stage in the lip print identification process. Research shows that the proposed method increased lip print identification accuracy for all tested template matching algorithms. After further improvements, our method can be used as a base for creating a personal identification system based on lip prints.

This paper presents a method of recognition of human lips. It can be treated as a new kind of biometric measure. During image preprocessing, the features are extracted from the lip print image. In the same step, image is denoised and normalized and ROI is determined. In the next stage, the normalized cross-correlation method was applied to estimation of the biometric parameters EER, FAR, and FRR. Investigations were conducted on 120 lip print images. These images come from University of Silesia public repository http://biometrics.us.edu.pl. The results obtained are very promising and suggest that the proposed recognition method can be introduced into professional forensic identification systems.

This paper describes the research aimed at finding the optimal configuration of the face recognition algorithm based on local texture descriptors (binary and ternary patterns). Since the identification module was supposed to be a part of the face tracking system developed for interactive wearable computer, proper feature selection, allowing for real-time operation, became particularly important. Our experiments showed that it is unfeasible to reduce the computational complexity of the process by choosing discriminant regions of interest on the basis of the training set. The application of simulated annealing, however, to the selection of the most discriminant LTP codes provided satisfactory results.

When dealing with large evolving datasets one needs machine learning models able to adapt to the growing number of information. In particular, stream classification is a research topic where classifiers need an ability to rapidly change their solutions and behave stably after many changes in training set structure. In this paper we show how recently proposed Extreme Entropy Machine can be trained in an online fashion supporting not only adding/removing points to/from the model but even changing the size of the internal representation on demand. In particular we show how one can build a well-conditioned covariance estimator in an online scenario. All these operations are guaranteed to converge to the optimal solutions given by their offline counterparts.

Adapting classification models to concept drift is one of the main challenges associated with applying these models in dynamic environments. In particular, the learned concept is not static and may change over time under the influence of varying conditions (i.e. varying context). Unlike existing approaches where only the most recent data are considered for adapting the model, we propose incorporating context awareness into the adaptation process. The goal is to utilise knowledge of relevant context variables to facilitate the selection of more relevant training data. Specifically, we propose to weight each training example based on the degree of similarity with the current context. To detect such similarity, we utilise two approaches: a simple difference between the context variable values and a distribution-based distance metric. The experimental analyses show that such explicit context utilisation results in a more effective data selection strategy and enables to produce more accurate predictions.

Societies around the world faced arrival of smart technologies in the last decade. Often interconnected, intelligent devices form new entity called Internet of Things (IoT). Mounted to commodities they are versatile tools for collecting various sorts of data about our behavior. Related applications require novel knowledge exploration methods handling large amount of observations containing complex data. Therefore, this paper introduces graph-stream structure as a capable tool for the complex process description. Further, it delivers a method for graph-stream processing making possible extraction of the compact ontological description of the recorded process. Introduced method uses novel online clustering algorithm and was verified experimentally on synthetic data sets.

Undersampling bagging ensembles specialized for class imbalanced data are considered. Particular attention is paid to Roughly Balanced Bagging, as it leads to better classification performance than other extensions of bagging. We experimentally analyze its properties with respect to bootstrap construction, deciding on the number of component classifiers, their diversity, and ability to deal with the most difficult types of the minority examples. We also discuss further extensions of undersampling bagging, where the data difficulty factors influence sampling examples into bootstraps.

This article proposes a new approach which helps to prevent the formation of self-crossings and loops in the parametric active contour model while tracking moving objects in video streams. The presented solutions mean that the process of tracking is stable in all subsequent video sequences. On the other hand, self-crossings of nodes and the resultant contour loops happen very often in the basic model and ruin the tracking process. The Gaussian filter was used to eliminate noise from video streams.

Due to requirements related to the maintenance of power transmission lines, it is necessary to diagnose their condition regularly. Among approaches being nowadays applied fundamental technical diagnostic methods are the vision inspections, often performed with use of cameras. The aim of this paper is to develop a method of automated recognition of insulators in images for the purposes of further computer analysis of their technical condition. Application of image segmentation by the statistical region merging method lets separate objects visible in images of very composed backgrounds. In order to recognize the insulators the template matching by an improved brute force method was used. The author proposes an approach which makes use of the fact that insulators are longitudinal and the problem with scaling and rotation variability is solved. The proposed algorithm can be applied to automatic recognition of insulators as well as any oblong elements in images.

Hands-free control of electronic devices gains increasing interest. The interaction based on the interpretation of hand gestures is convenient for users. However, it requires adequate techniques to capture user movement and appropriate onscreen interface. The hand movements in touchless graphical user interface are translated into the motion of a pointer on a display. The main question is how to convert hand gestures into casual and comfortable text entry. The paper focuses on the evaluation of text input techniques in a touchless interface. Well-known traditional solutions and some innovations for text input have been adapted to noncontact onscreen keyboard interface and subjected to examination. The examined solutions include: single hand and double hands QWERTY-based virtual keyboard, swipe text input, and the 8pen (The 8pen solution, http://www.8pen.com/ [1]) based technique.

This paper presents an outcome of experiments on a self-localization system for small mobile robots that involves the use of QR codes as artificial landmarks. The QR codes are employed for the double purpose of localization marks and data carriers for navigation-relevant informations. As we want to make our robots autonomous and independent from external cameras for localization, we investigate the use of an omnidirectional vision system. Then, we demonstrate that a hybrid solution consisting of an omnidirectional camera with a mirror and a classic front-view camera provides better localization results and application flexibility than either of these configurations applied alone. The hybrid vision system is inspired by the peripheral and foveal vision cooperation in animals. We demonstrate that the omnidirectional camera enables the robot to detect quickly landmark candidates and track the already recognized QR codes, while the front-view camera guided by the omnidirectional information enables precise measurements of the landmark position over extended distances and reading of the extra data carried by the QR code.

The problem of light microscope images enhancement by filtering for recognition pathologic liver tissues in fish is considered in the paper. The problem follows from the necessity of monitoring the sea water pollutions caused by mercury compounds and their influence on living organisms. It is proposed to use image filtering based on morphological spectra to enhance visibility of liver lesions in the images in order to extract morphologic-statistical parameters useful in automatic tissues classification into normal and pathologic classes. It is shown that selected components of the 4th range morphologic spectra (MS4) are the most suitable to discriminate normal and pathologic liver tissues. The selected spectral components are characterized by their estimated mean values, standard deviations and kurtoses. The so-obtained morphologic-statistical parameters have been used to construct the learning sets for two types of image classifiers: based on the nearest mean and knearest neighbors rules. It is shown that preliminary image filtering by morphological spectra-based filters improves spatial distribution of the recognized normal and pathologic objects in the parameter space.

Data obtained by hyperspectral imaging gives us enough information to recreate the human vision, and also to extend it by a new methods to extract features coded in a light spectra. This work proposes a set of functions, based on abstraction of natural photoreceptors. The proposed method was employed as the feature extraction for the classification system based on combined approach and compared with other state-of-art methods on the basis of the selected benchmark images.

This paper presents a visual system for real-time eye detection and tracking in the near-infrared (NIR) video streams for drivers’ monitoring. The system starts with crude eye position estimation based on an eye model suitable for NIR processing. In the next step, eye regions are verified with the classifier operating in the higher-order decomposition of the tensor of eye prototypes. Finally, the process is augmented with the linear tracker which facilitates eye detection and allows real-time operation necessary in the automotive environment. The reported experiments show high accuracy and real-time operation of the system in the car.

This paper deals with the problem of estimating the similarity of clothing for the purpose of fashion-related e-commerce systems. The images presenting fashion models are segmented and analyzed in order to detect clothing characteristics. We propose a method based on human pose estimation and body parts segmentation, followed by the analysis of dominant color and structural similarity, independently for particular body segments. The algorithm can be utilized to perform clusterization or in the simpler case—to directly search for similar outfits. The experiments performed using 1800 real-life photos proved the applicability of the proposed approach.

This paper addresses the problem of processing data from road cameras for providing work parameters for traffic flow control systems. 3D wavelet transformation of image sequences is proposed to represent the traffic. In order to reduce the sensitivity to ambient light changes, of the road scene, a linear function of the coefficients is used to represent the traffic flow. The parameters of the linear function are determined using real traffic data by minimizing the MSE of vehicle detection functions. The developed algorithm was tested using a database of image sequences. Test results prove that it can be applied to determine traffic flow values for control systems. Instead of the usual elaborate image sequence processing, a hardware-based 3D wavelet transformation may be added to the control system.

In previous papers we presented a computer system to detect erosions and osteophytes from hand radiographs (the most common symptoms of rheumatic diseases) based on the shape analysis of the joint surfaces borders. Such borders are obtained automatically using algorithms which were also proposed in our previous articles. In this paper, we consider a new approach which analyzes patches located at the joint surfaces borders in order to determine which of them correspond to the lesions. Vectors of features which are used to classify patches are calculated by applying Schmid filter with various frequencies and scales. Additional features are obtained using inpainting. Vectors are analyzed based on Gaussian mixture model calculated with expectation maximization algorithm. The accuracy is measured with area under curve of the receiver-operating characteristic. The conducted experiments proved that, the shape approach described in our previous work can be improved by applying Schmid filter and the inpainting approach in the parsing stage, especially, in case of the lower MCP and upper PIP surfaces for which classification still remains inaccurate.

In this paper, a parametric histogram-based image segmentation method is used where the gray level histogram is considered as a finite mixture of asymmetric generalized Gaussian distribution (AGGD). The choice of AGGD is motivated by its flexibility to adapt the shape of the data including the asymmetry. Here, the method of moment estimation combined to the expectation–maximization algorithm (MME/EM) is originally used to estimate the mixture parameters. The proposed image segmentation approach is achieved in radiographic imaging where the image often presents an histogram with a complex shape. The experimental results provided in terms of histogram fitting error and region uniformity measure are comparable to those of the maximum likelihood method (MLE/EM) with the advantage that MME/EM method reveals to be more robust to the EM initialization than MLE/EM.

Correct and accurate classification of ECG patterns in a long-term record requires optimal selection of feature vector. We propose a machine learning algorithm that learns from short randomly selected signal strips and, having an approval from a human operator, classifies all remaining patterns. We applied a genetic algorithm with aggressive mutation to select few most distinctive features of ECG signal. When applied to the MIT-BIH Arrhythmia Database records, the algorithm reduced the initial feature space of 57 elements to 3–5 features optimized for a particular subject. We also observe a significant reduction of misclassified beats percentage (from 2.7 % to 0.7 % in average for SVM classifier and three features) with regard to automatic correlation-based selection.

Environmental Microbiology (EM) is an importantShirahama, K.scientificLi, C. field, whichBoukhers, Z.investigatesZou, Y. the ecologicalJiang, T. usage of differentGrzegorzek, M. microorganisms. Traditionally, researchers look for the information of microorganisms by checking references or consulting experts. However, these methods are time-consuming and not effective. To increase the effectiveness of EM information search, we propose a novel approach to aid the information searching work using Content-based Image Retrieval (CBIR). First, we use an microorganism image as input data. Second, image segmentation technique is applied to obtain the shape of the microorganism. Third, we extract shape feature from the segmented shape to represent the microorganism. Especially, we use a contour-based shape feature called Internal Structure Histogram (ISH) to describe the shape, which can use angles defined on the shape contour to build up a histogram and represent the structure of the microorganism. Finally, we use Euclidean distances between each ISHs to measure the similarity of different EM images in the retrieval task, and use Average Precision (AP) to evaluate the retrieval result. The experimental result shows the effectiveness and potential of our EM-CBIR system.

This paper presents the concept of kinematic control of prosthetic hand with 13 d.o.f., while grasping objects of different shapes and sizes. The concept refers to the process of healthy hand motion control performed by the human nervous system. Planning of grip is based on kinematic model of the hand. The parameters of subsequent phases of the gripping process were determined experimentally from the measurements carried out with a laboratory model of hand. It was assumed that the final arrangement of fingers on the workpiece is determined on the basis of information from the touch sensor system.

This paper deals with the problem of bioprosthetic hand control via recognition of user intent on the basis of electromyography (EMG) and mechanomyography (MMG) signals acquired from the surface of a forearm. As a method of signal parameterization the Hilbert–Huang (HH) transform is applied which is an effective tool for reduction of feature space dimension. The performance of proposed recognition method based on HH transform of EMG and MMG signals was experimentally compared against an autoregressive model of dimensionality reduction using real data concerning the recognition of five types of grasping movements. The experimental results show that the HH transform approach with root mean square of amplitude feature outperforms an autoregressive method.

Two techniques of the image analysis of Papanicolaou stains are compared in this paper—standard deviation and standard deviation ratio for cell nuclei. The image analysis is based on diagonal details obtained from multiresolutional analysis using wavelets. Two best wavelets are presented ‘coif2’ and ‘sym1.’ Results show the importance of standard deviation ratios and smallest diagonal details for classification of cell together with cell nucleus area. Classification of cells allows rapid discrimination of cells for further analysis of them by cytoscreener.

This paper reports the results of the experiment addressing the recovery from interruption phenomenon in terms of brain activity patterns. The aim of the experiment was to find out whether it is possible to find any significant differences in brain activity between subjects performing the task in the recovery period better or worse than the control group. The main outcome from the experiment was that the brain activity of the subjects who performed better than the control group did not change significantly during back to task period compared to interruption period. On the contrary, for subjects whose performance was worse than in the control group, the significant changes in signal power in some frequency bands were found.

Electrical load forecasting is an important problem concerning safe and cost-efficient operation of the power system. Although many techniques are used to predict an electrical load, a research into constructing more accurate methods and software tools is still being conducted over the world. In this paper an experimental application for improving an accuracy of an electrical load prediction is presented. It is based on the syntactic pattern recognition approach and FGDPLL(k) string automata. The application has been tested on the real data delivered by one of the Polish electrical distribution companies.

We present the METINUSBosch, R. (METhod of ImmunohistochemicalRoszkowiak, L. NUclei Segmentation)Korzynska, A., which is a improvedLejeune, M. and modifiedLopez, C. version of supporting tool for pathologists from 2010. The method supports examination of immunohistochemically stained thin tissue sections from biopsy of follicular lymphoma patients. The software localizes and counts FOXP3 expression in the cells’ nuclei supporting standard procedure of diagnosis and prognosis. The algorithm performs colour separation followed by object extraction and validation. Objects with statistical parameters not in specified range are disqualified from further assessment. To calculate the statistics we use the following: three channels of RGB, three channels of Lab colour space, brown channel and three layers completed with colour deconvolution. Division of the objects is done with support of watershed and colour deconvolution algorithm. Evaluation was performed on arbitrarily chosen 20 images with moderate quality of most typical tissues. We compared results of improved method with the previous version in the context of semiautomatic, pathologist controlled, computer-aided result of quantification as reference. Comparison is based on quantity of nuclei located per image using Kendall’s tau-b correlation coefficient. It shows concordance of 0.91 between results of proposed method and reference, while with previous version it is only 0.71.

Swipe typing has been designed for touchscreen devices and consists of sliding a finger or stylus through consecutive letters lifting only between words. Since the tracked path contains many redundant letters the accurate recognition of intended word requires a good input path analyser and a word search engine. There are many proprietary solutions of swipe typing available on the market; however, all of them focus on touchscreen devices. On the other hand, there is a growing number of non-contact interfaces operated with gestures. These interfaces limit their operation to pointer manipulation in graphical user interface. The problem of typing in most cases is omitted. Of course touchless interfaces are not designed for text entry purposes but at least some decent possibility of text entry is necessary. In this paper the word recognition algorithm from the tracked path created by the hand movement in front of the wall screen projector is proposed. It is compared with other solutions based on individual key selection in touchless environment.

Defects in furniture elements were detected using data from a commercially available structured light 3D scanner. Out-of-plane deviations down to 0.15 mm were analyzed successfully. The hierarchical, iterated version of the Hough transform was used. The calculation of position of the plane could be separated from that of its direction due to the assumption of nearly horizontal location of the plane, which is natural when the tested elements lie on a horizontal surface.

K-means clustering algorithm has been used to classify patterns of Japanese candlesticks which accompany the approach to trend reversals in the prices of several assets registered in the Warsaw stock exchange (GPW). It has been found that the trend reversals seem to be preceded by specific combinations of candlesticks with notable frequency. Surprisingly, the same patterns appear in both “bullish” and “bearish” trend reversals. The above findings should stimulate further studies on the problem of applicability of the so-called technical analysis in the stock markets.

Line following robots could be applied in numerous applications with artificial or natural line. The proposed algorithm uses tracklets and Cartesian-to-polar conversion together with Viterbi algorithm for the estimation of line. The line could be low contrast or deteriorated and Monte Carlo tests are applied for the analysis of algorithm properties. Two algorithms are presented and compared—Viterbi and proposed Tracklet-based Viterbi Track-Before-Detect algorithms. Both of them are evaluated and properties are presented. The proposed algorithm could be better for smoother lines if higher noise disturb images.

The classification tries to assign the best category to given unknown records based on previous observations. It is clear that with the growing amount of data, any classification algorithm can be very slow. The learning speed of many developed state-of-the-art algorithms like deep neural networks or support vector machines is very low. Evolutionary-based approaches in classification have the same problem. This paper describes five different evolutionary-based approaches that solve the classification problem and run in real time. This was achieved by using GPU parallelization. These classifiers are evaluated on two collections that contains millions of records. The proposed parallel approach is much faster and preserve the same precision as a serial version.

In this paper, an efficient hardware architecture, enabling the computation of LBP-based image region descriptors is presented. The complete region descriptor is formed by combining individual local descriptors and arranging them into a grid, as typically used in object detection and recognition. The proposed solution performs massively parallel, pipelined computations, facilitating the processing of over two hundred VGA frames per second, and can easily be adopted to different window and grid sizes for the use of other descriptors.

The paper proposes the $$\textsf {Unified\;Process\;Management\;for\;Service\;and\;Manufacture\;(UPM}^{\mathsf {Srv}}_{\mathsf {Mnf}}\textsf {)}$$UnifiedProcessManagementforServiceandManufacture(UPMMnfSrv), the intent of which is to constitute integrated platform for modelling and exchange of information in the field of production processes. The $${\textsf {UPM}}^{\mathsf {Srv}}_{\mathsf {Mnf}}$$UPMMnfSrv consists of a theoretical basis, formal notation, modeling language, algorithms, and methods of processing information as well as modeling methodology. This article focuses on the passage concerning the most theoretical foundations in the identification of selected categories and relations such as generalization-specialization, part–whole, class-feature, class-instance, and property-instance. The platform used to build $${\textsf {UPM}}^{\mathsf {Srv}}_{\mathsf {Mnf}}$$UPMMnfSrv is the Semantic Knowledge Base (SKB), which is the result of a research project on methods of representation and processing of knowledge.

This paper presents method of image conversion into point image representation. Conversion is carried out with the use of small image gradients. Layout of binary values of point image representation is in accordance with the edges of objects comprised in the source image. Vehicles are detected through analysis of the detection field state. The state of the detection field is determined on the basis of the sum of the edge points within the detection field. The proposed method of vehicle detection is fast and simple computationally. Vehicle detection with the use of image conversion into point image representation is efficient and can by used in real-time processing. Experimental results are provided.

Control of dental millingWozniak, M.processesJackowski, K. is a taskCorchado, E. which can significantlyQuintián, H. reduce productionJankowski, D. costs due to possible savings in time. Appropriate setup of production parameters can be done in a course of optimisation aiming at minimising selected objective function, e.g. time. Nonetheless, the main obstacle here is lack of explicitly defined objective functions, while model of relationship between the parameters and outputs (such as costs or time) is not known. Therefore, the model must be discovered in advance to use it for optimisation. Machine learning algorithms serve this purpose perfectly. There are plethoras of competing methods and the question is which shall be selected. In this paper, we present results of extensive investigation on this question. We evaluated several well-known classical regression algorithms, ensemble approaches and feature selection techniques in order to find the best model for dental milling model.

This paper presents the hybrid approach using fuzzy logic and rough sets used as a pattern recognition framework. Both fuzzy and rough sets have been introduced to deal with vagueness and uncertain data in artificial intelligence applications. In general, fuzzy logic can be related to vagueness, while rough sets deal with indiscernibility. In our work we propose two-stage algorithm. At the first stage, an optimization procedure is applied to reduce the number of features for fuzzy membership functions and to find the optimal granulation for rough sets, respectively. In the second stage, two-step classifier is used. We tested our attempt using Handprinted Forms and Characters Database containing the full page binary images of 3699 handwriting sample forms. For any segmented image classification, a crucial part lies in the proper feature extraction method. In our work, cross corner feature algorithm was used as a main tool.

The article presents an approach to environmental sound recognition that uses selected methods from the field of digital image processing and recognition. The proposed technique adopts the assumption that an audio signal can be converted into a visual representation, and processed further, as an image. At the first stage the audio data are converted into rectangular matrices called feature maps. Then a two-step approach is applied: the construction of a representative database of reference samples and the identification of test samples. The process of building the database employs two-dimensional linear discriminant analysis. Then the recognition operation is carried out in a reduced feature space that has been obtained by two-dimensional Karhunen–Loeve projection. At the classification stage, a minimum distance classifier is applied to different features. As it is shown, the results are very encouraging and can be a base for many practical audio applications.

The article addresses a problem of audio signal classification employing image processing and recognition methods. In such an approach, vectorized audio signal features are converted into a matrix representation (feature map), and then processed, as a regular image. In the paper, we present a process of creating a low-dimensional feature space by means of two-dimensional Linear Discriminant Analysis and projecting input feature maps into this subspace using two-dimensional Karhunen–Loeve Transform. The classification is performed in the reduced feature space by means of voting on selected distance metrics applied for various features. The experiments were aimed at finding an optimal (in terms of classification accuracy) combination of six feature types and five distance metrics.The found combination makes it possible to perform audio classification with high accuracy, yet the dimensionality of resulting feature space is significantly lower than input data.

The paper presents a method which allows to improve tracking abilities of conventional background subtraction-based algorithm. The presented algorithm which is a result of the studies is a hybrid method consisting of the Kernel Density Estimation (KDE) background subtraction tracking method and the Edge Histograms Displacement Calculation (EHDC) algorithm. Tracking ratios before and after merging with EHDC have been measured and presented. The paper also describes an algorithm eliminating cyclic changes in image’s intensities values, which have significant influence on the input data for the hybrid algorithm. The influence of moving-camera video specificity on the output data has been pointed out.

The web has developed into one of the most relevant data sources and becomes now a broad knowledge base for almost all fields. Its content grows faster, and its size becomes larger every day. Due to this big amount of data, web page classification becomes crucial since users encounter difficulties in finding what they are seeking, even though they use search engines. Web page classification is the process of assigning a web page to one or more classes based on previously seen labeled examples. Web pages contain a lot of contextual features that can be used to enhance the classification’s accuracy. In this paper, we present a similarity computation technique that is based on implicit links extracted from the query-log, and used with K-Nearest Neighbors (KNN) in web page classification. We also introduce an implicit links-based probability computation method used with Naive Bayes (NB) for web page classification. The new computed similarity and probability help enrich KNN and NB respectively for web page classification. Experiments are conducted on two subsets of Open Directory Project (ODP). Results show that: (1) when applied as a similarity for KNN, the implicit links-based similarity helps improve results. (2) the implicit links-based probability helps ameliorateBenkhalifa M. results provided by NB using only text-based probability.

Currently, the growing popularity of publicly available web services is one of the driving forces for so-called “web hacking” activities. The main contribution of this paper is the semi-unsupervised anomaly detection method for HTTP traffic anomaly detection. We made the assumption that during the learning phase (for the captured volume of HTTP traffic), only small friction of samples is labelled. Our experiments show that the proposed method allows us to achieve the ratios of true positive and false positive errors below 1 %.

This paper dealsPovh, J. with automatic two classŽnidaršič, M.document-levelBučar, J. sentiment classification. We retrieved textual documents with political, business, economic and financial content from five Slovenian web media. By annotating a sample of 10,427 documents, we obtained a labelled corpus in the Slovenian language. Five classifiers were evaluated on this corpus: multinomial naïve Bayes, support vector machines, random forest, k-nearest neighbour and naïve Bayes, out of which the first three were used also in the assessment of the pre-processing options. Among the selected classifiers, multinomial naïve Bayes outperforms the naïve Bayes, k-nearest neighbour, random forest and support vector machines classifier in terms of classification accuracy. The best selection of pre-processing options achieves more than 95 % classification accuracy with Naïve Bayes Multinomial and more than 85 % with support vector machines and random forest classifier.

Patients with chronic diseases are increasing around the globe. Healthcare professionals attempt to find possible causes of the chronic diseases. One of the most possible causes is the sleep disorder. Sleep apnea, OSA and CSA, may be an evidence of chronic diseases. In order to detect the sleep apnea, the polysomnography (PSG) or the sleep test is required for patients. A number of parameters will be collected on patients whilst they are asleep. However, due to the limitation of the PSG test in some countries, researchers attempt to find other available alternative approaches. In this research work, a mobile application has been constructed to perform a screening test of OSA. With our initial experiment test, 74.70 % instances have been correctly classified. An application of SMOTE into a minority class is performed and achieves up to 80.10 % correctly classified instances. Limitations of the mobile application and our technique have also discussed.

In this article we present a network traffic DDoS attacks detection method based on modeling the variability with the use of conditional average and variance in examined time series. Variability predictions of the analyzed network traffic are realized by estimated statistical models ARFIMA and FIGARCH. We propose simple parameter estimation models with the use of maximum likelihood function. The choice of sparingly parameterized form of the models is realized by means of information criteria representing a compromise between brevity of representation and the size of the prediction error. In the described method we propose using statistical relations between predicted and analyzed network traffic in order to detect abnormal behavior possibly being a result of a network attack. Performed experiments confirmed effectiveness of the analyzed method and cogency of the statistical models. abstract environment.

The analysis of infrared (IR) images obtained from a robot-mounted camera is presented, with the purpose to reconstruct the 3D surface of texture-less objects located in close range to the camera. The prospective application of this approach is object’s pose recognition with the aim of object grasping by a robot hand. Algorithms are developed that rely on the analysis of the luminance distribution in the IR image (the so-called shape-from-shading approach), followed by a depth-map approximation. Laboratory tests were carried out in order to evaluate the quality of obtained depth maps on some reference objects and to compare the estimated depth maps with corresponding point clouds acquired by a MS-Kinect device.

Advent of RGB-D sensors fostered the progress of computer vision algorithms spanning from object recognition, object and scene modelling to human activity recognition. This paper presents a new flavour of ICP algorithm developed for the purpose of pair-wise registration of colour point clouds generated from RGB-D images. After a brief introduction to the registration problem, we analyze the ICP algorithm and survey its different flavours in order to indicate potential methods of injecting of colour into it. Our consideration led to a solution, which we validate experimentally on colour point clouds from the publicly available dataset.

In this paper the GPU-accelerated range sensors simulation is discussed. Range sensors generate large amount of data per second and to simulate these high-performance simulation is needed. We propose to use parallel ray tracing on graphics processing units to improve the performance of range sensors simulation. The multiple range sensors are described and simulated using NVIDIA OptiX ray tracing engine. This work is focused on the performance of the GPU acceleration of range images simulation in complex environments. Proposed method is tested using several state-of-the-art ray tracing datasets. The software is publicly available as an open-source project SensorSimRT.

Recent advancements in 3D video generation, processing, compression, and rendering increase accessibility to 3D video content. However, the majority of 3D displays available on the market belong to the stereoscopic display class and require users to wear special glasses in order to perceive depth. As an alternative, autostereoscopic displays can render multiple views without any additional equipment. The depth perception on stereoscopic and autostereoscopic displays is realized via a binocular depth cue called stereopsis. Another important depth cue, that is not exploited by autostereoscopic displays, is motion parallax which is a monocular depth cue. To enable the motion parallax effect on a 2D display, we propose to use the Kinect sensor to estimate the pose of the viewer. Based on pose of the viewer the real-time view synthesis software adjusts the view and creates the motion parallax effect on a 2D display. We believe that the proposed solution can enhance the content displayed on digital signature displays, kiosks, and other advertisement media where many users observe the content during move and use of the glasses-based 3D displays is not possible or too expensive.