Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications

Pattern recognition systems have been increasingly used in security applications, although it is known that carefully crafted attacks can compromise their security. We advocate that simulating a proactive arms race is crucial to identify the most relevant vulnerabilities of pattern recognition systems, and to develop countermeasures in advance, thus improving system security. We summarize a framework we recently proposed for designing proactive secure pattern recognition systems and review its application to assess the security of biometric recognition systems against poisoning attacks.

The ensemble classification paradigm is an effective way to improve the performance and stability of individual predictors. Many ways to build ensembles have been proposed so far, most notably bagging and boosting based techniques. Evolutionary algorithms (EAs) also have been widely used to generate ensembles. In the context of heterogeneous ensembles EAs have been successfully used to adjust weights of base classifiers or to select ensemble members. Usually, a weighted sum is used for combining classifiers outputs in both classical and evolutionary approaches. This study proposes a novel genetic program that learns a fusion function for combining heterogeneous-classifiers outputs. It evolves a population of fusion functions in order to maximize the classification accuracy. Highly non-linear functions are obtained with the proposed method, subsuming the existing weighted-sum formulations. Experimental results show the effectiveness of the proposed approach, which can be used not only with heterogeneous classifiers but also with homogeneous-classifiers and under bagging/boosting based formulations.

A reduction operator transforms a binary picture only by changing some black points to white ones, which is referred to as deletion. Sequential reductions may delete just one point at a time, while parallel reductions can alter a set of points simultaneously. Two reductions are called equivalent if they produce the same result for each input picture. This work lays a bridge between the parallel and the sequential strategies. A class of deletion rules are proposed that provide 2D parallel reductions being equivalent to sequential reductions. Some new sufficient conditions for topology-preserving parallel reductions are also reported.

A 3D object decomposition method is presented, based on the polygonal approximation of the distance labeled curve skeleton. Polygonal approximation is accomplished to divide each skeleton branch into a number of segments along which no significant changes exist as regards curvature or distance label. Each segment is interpreted as the spine of a simple region, which is characterized by i) absence of significant curvature changes along its boundary and ii) thickness that is either constant or evolves linearly along the region. Quantitative information on shape, size, position and orientation of a simple region can be easily derived from spatial coordinates and distance labels of the extremes of the associated spine. Simple regions associated to spines sharing a common extreme partially overlap with each other. Object decomposition into disjoint regions is obtained by suitably dividing each overlapping region among the simple regions including it.

We present a new approach to image analysis in temporal sequence of images (data cube). Our method is based on high-order statistical moments (skewness and kurtosis) giving interesting information about a dynamic event in the temporal sequence. The moments enable precise determination of the ”turning points” in the temporal sequence of images. The moment’s curves are analyzed by continuous complex Morlet wavelet that leads to the description of quasi-periodic processes in the investigated event as a time sequence of local spectra. These local spectra are compared with Fourier spectrum. We experimentally illustrate the performance on the real data from astronomical observations.

The Distance Transform (DT) is one of the classical operators in image processing, and can be used in Pattern Recognition and Data Mining, and there is currently a great demand for efficient parallel implementations on graphics cards, known as GPU. This paper presents simple and effective ways to implement the DT using decompositions of erosions with structuring functions implemented on GPU. The DT is equivalent to a morphological erosion of the binary image by a specific structuring function. However, this erosion can be decomposed by a sequence of erosions using small structuring functions. Classical and efficient algorithms of the DT are implemented on CPU. New 1D and 2D algorithms are implemented on GPU, using decomposition of structuring functions, inspired by implementations of convolution filters. All the GPU implementations used in this paper are known as

brute-force

, and even then present excellent results, comparable to the best CPU algorithms, which might contribute to future applications in image processing.

Single-Gaussian and Gaussian-Mixture Models are utilized in various pattern recognition tasks. The model parameters are estimated usually via Maximum Likelihood Estimation (MLE) with respect to available training data. However, if only small amount of training data is available, the resulting model will not generalize well. Loosely speaking, classification performance given an unseen test set may be poor. In this paper, we propose a novel estimation technique of the model variances. Once the variances were estimated using MLE, they are multiplied by a scaling factor, which reflects the amount of uncertainty present in the limited sample set. The optimal value of the scaling factor is based on the Kullback-Leibler criterion and on the assumption that the training and test sets are sampled from the same source distribution. In addition, in the case of GMM, the proper number of components can be determined.

One of the key steps in 3D object recognition is the matching between an input cloud and a cloud in a database of known objects. This is usually done using a distance function between sets of descriptors. In this paper we propose to study how several distance functions (some already available and other new proposals) behave experimentally using a large freely available household object database containing 1421 point clouds from 48 objects and 10 categories. We present experiments illustrating the accuracy of the distances both for object and category recognition and find that simple distances give competitive results both in terms of accuracy and speed.

In recent years, Evolutionary Algorithms (EAs) have been remarkably useful to improve the robustness of Artificial Neural Networks (ANNs). This study introduces an experimental analysis using an EAs aimed to evolve ANNs architectures (the FS-EPNet algorithm) to understand how neural networks are evolved with a steady-state algorithm and compare the Single-step-ahead (SSP) and Multiple-step-ahead (MSP) methods for prediction tasks over two test sets. It was decided to test an inside-set during evolution and an outside-set after the whole evolutionary process has been completed to validate the generalization performance with the same method (SSP or MSP). Thus, the networks may not be correctly evaluated (misleading fitness) if the single SSP is used during evolution (inside-set) and then the MSP at the end of it (outside-set). The results show that the same prediction method should be used in both evaluation sets providing smaller errors on average.

This work presents a new method to apply the Hough Transform to 2D and 3D cloud points using the conformal geometric algebra framework. The objective is to detect geometric entities, with the use of simple parametric equations and the properties of the geometric algebra. We show with real images and RGB-D data that this new method is very useful to detect lines and circles in 2D and planes and spheres in 3D.

Recently, the sieve bootstrap method has been successfully used in prediction of nonlinear time series. In this work we study the performance of the prediction intervals based on the sieve bootstrap technique, which does not require the distributional assumption of normality as most techniques that are found in the literature. The construction of prediction intervals in the presence of patchy outliers are not robust from a distributional point of view, leading to an undesirable increase in the length of the prediction intervals.

In the analysis of financial time series it is common to have irregular observations that have different types of outliers, isolated and in groups. For this reason we propose the construction of prediction intervals for returns based in the winsorized residual and bootstrap techniques for financial time series. We propose a novel, simple, efficient and distribution free resampling technique for developing robust prediction intervals for returns and volatilities for TGARCH models. The proposed procedure is illustrated by an application to known synthetic time series.

In this paper we propose a new descriptor for 3D point clouds that is fast when compared to others with similar performance and its parameters are set using a genetic algorithm. The idea is to obtain a descriptor that can be used in simple computational devices, that have no GPUs or high computational capabilities and also avoid the usual time-consuming task of determining the optimal parameters for the descriptor. Our proposal is compared with other similar algorithms in a public available point cloud library (PCL [1]). We perform a comparative evaluation on 3D point clouds using both the object and category recognition performance. Our proposal presents a comparable performance with other similar algorithms but is much faster and requires less disk space.

Enumeration and reconstruction of certain types of polyominoes, according to several parameters, are frequently studied problems in combinatorial image processing. Polyominoes with fixed projections play an important role in discrete tomography. In this paper, we provide a linear-time algorithm for reconstructing

hv

-convex polyominoes with minimal number of columns satisfying a given horizontal projection. The method can be easily modified to get solutions with any given number of columns. We also describe a direct formula for calculating the number of solutions with any number of columns, and a recursive formula for fixed number of columns.

A new constraint acquisition method for parwise-constrained data clustering based on user-feedback is proposed. The method searches for non-redundant intra-cluster and inter-cluster query-candidates, ranks the candidates by decreasing order of interest and, finally, prompts the user the most relevant query-candidates. A comparison between using the original data representation and using a learned representation (obtained from the combination of the pairwise constraints and the original data representation) is also performed. Experimental results shown that the proposed constraint acquisition method and the data representation learning methodology lead to clustering performance improvements.

Linear or non-linear data transformations are widely used processing techniques in clustering. Usually, they are beneficial to enhancing data representation. However, if data have a complex structure, these techniques would be unsatisfying for clustering. In this paper, based on the auto-encoder network, which can learn a highly non-linear mapping function, we propose a new clustering method. Via simultaneously considering data reconstruction and compactness, our method can obtain stable and effective clustering. Experiments on three databases show that the proposed clustering model achieves excellent performance in terms of both accuracy and normalized mutual information.

The Mahalanobis distance (MD) is a widely used measure in Statistics and Pattern Recognition. Interestingly, assuming that the data are generated from a Gaussian distribution, it considers the covariance matrix to evaluate the distance between a data point and the distribution mean. In this work, we generalize MD for distributions in the exponential family, providing both, a definition in terms of the data density function and a computable version. We show its performance on several artificial and real data scenarios.

Encoding an object essence in terms of self-similarities between its parts is becoming a popular strategy in Computer Vision. In this paper, a new similarity-based descriptor, dubbed Structural Similarity Cross-Covariance Tensor is proposed, aimed to encode relations among different regions of an image in terms of cross-covariance matrices. The latter are calculated between low-level feature vectors extracted from pairs of regions. The new descriptor retains the advantages of the widely used covariance matrix descriptors [1], extending their expressiveness from local similarities inside a region to structural similarities across multiple regions. The new descriptor, applied on top of HOG, is tested on object and scene classification tasks with three datasets. The proposed method always outclasses baseline HOG and yields significant improvement over a recently proposed self-similarity descriptor in the two most challenging datasets.

In this paper, we propose a novel satisfaction mechanism, named “Dynamic

K

”, which could be introduced in any Class Association Rules (CAR) based classifier, to determine the class of unseen transactions. Experiments over several datasets show that the new satisfaction mechanism has better performance than the main satisfaction mechanism reported (“Best Rule”, “Best

K

Rules” and “All Rules”). Additionally, the experiments show that “Dynamic

K

” obtains the best results independent of the CAR-based classifier used.

A weighted naïve Bayes classifier using Renyi entropy is proposed. Such a weighted naïve Bayes classifier has been studied so far, aiming at improving the prediction performance or at reducing the number of features. Among those studies, weighting with Shannon entropy has succeeded in improving the performance. However, the reasons of the success was not well revealed. In this paper, the original classifier is extended using Renyi entropy with parameter

α

. The classifier includes the regular naïve Bayes classifier in one end (

α

 = 0.0) and naïve Bayes classifier weighted by the marginal Bayes errors in the other end (

α

 = ∞). The optimal setting of

α

has been discussed analytically and experimentally.

The evaluation of clustering algorithms is a field of Pattern Recognition still open to extensive debate. Most quality measures found in the literature have been conceived to evaluate non-overlapping clusterings, even when most real-life problems are better modeled using overlapping clustering algorithms. A number of desirable conditions to be satisfied by quality measures used to evaluate clustering algorithms have been proposed, but measures fulfilling all conditions still fail to adequately handle several phenomena arising in overlapping clustering. In this paper, we focus on a particular case of such desirable conditions, which existing measures that fulfill previously enunciated conditions fail to satisfy. We propose a new evaluation measure that correctly handles the studied phenomenon for the case of overlapping clusterings, while still satisfying the previously existing conditions.

The notion of homogeneous logical proportions has been recently introduced in close relation with the idea of analogical proportion. The four homogeneous proportions have intuitive meanings, which can be related with classification tasks. In this paper, we proposed a supervised classification algorithm using homogeneous logical proportions and provide results for all. A final comparison with previous works using similar methodologies and with other classifiers is provided.

Precise segmentation of bone cancer is an important step for several applications. However, the achievement of this task has proven problematic due to lack of contrast and the non homogeneous intensities in many modalities such as MRI and CT-scans. In this paper we investigate this line of research by introducing a new method for segmenting bone cancer. Our segmentation process involves different steps: a registration step of different image modalities, a fuzzy-possibilistic classification (FPCM) step and a final segmentation step based on a variational model. The registration and the FPCM algorithms are used to locate and to initialize accurately the deformable model that will evolve smoothly to delineate the expected tumor boundaries. Preliminary results show accurate and promising detection of the cancer region.

The text below describes a short introduction to extreme learning machines (ELM) enlightened by new developed applications. It also includes an introduction to deep belief networks (DBN), noticeably tuned into the pattern recognition problems. Essentially, the deep belief networks learn to extract invariant characteristics of an object or, in other words, an DBN shows the ability to simulate how the brain recognizes patterns by the contrastive divergence algorithm. Moreover, it contains a strategy based on both the kernel (and neural) extreme learning of the deep features. Finally, it shows that the DBN-ELM recognition rate is competitive (and often better) than other successful approaches in well-known benchmarks. The results also show that the method is extremely fast when the neural based ELM is used.

Spectral clustering techniques have shown their capability to identify the data relationships using graph analysis, achieving better accuracy than traditional algorithms as

k

-means. Here, we propose a methodology to build automatically a graph representation over the input data for spectral clustering based approaches by taking into account the local and global sample structure. Regarding this, both the Euclidean and the geodesic distances are used to identify the main relationships between a given point and neighboring samples around it. Then, given the information about the local data structure, we estimate an affinity matrix by means of Gaussian kernel. Synthetic and real-world datasets are tested. Attained results show how our approach outperforms, in most of the cases, benchmark methods.

In this work we present a novel approach to transfer knowledge between reinforcement learning tasks with continuous states and actions, where the transition and policy functions are approximated by Gaussian Processes (GPs). The novelty in the proposed approach lies in the idea of transferring qualitative knowledge between tasks, we do so by using the GPs’ hyper-parameters used to represent the state transition function in the source task, which represents qualitative knowledge about the type of transition function that the target task might have. We show that the proposed technique constrains the search space, which accelerates the learning process. We performed experiments varying the relevance of transferring the hyper-parameters from the source task into the target task and show, in general, a clear improvement in the overall performance of the system when compared to a state of the art reinforcement learning algorithm for continuous state and action spaces without transfer.

Optimal Subsequence Bijection (OSB) is a method that allows comparing two sequences of endnodes of two skeleton graphs which represent articulated shapes of 2D images. The OSB dissimilarity function uses a constant penalty cost for all endnodes not matching between two skeleton graphs; this can be a problem, especially in those cases where there is a big amount of not matching endnodes. In this paper, a new penalty scheme for OSB, assigning variable penalties on endnodes not matching between two skeleton graphs, is proposed. The experimental results show that the new penalty scheme improves the results on supervised classification, compared with the original OSB.

Missing values can occur frequently in many real world situations. Such is the case of multi-way data applications, where objects are usually represented by arrays of 2 or more dimensions e.g. biomedical signals that can be represented as time-frequency matrices. This lack of attributes tends to influence the analysis of the data. In classification tasks for example, the performance of classifiers is usually deteriorated. Therefore, it is necessary to address this problem before classifiers are built. Although the absence of values is common in these types of data sets, there are just a few studies to tackle this problem for classification purposes. In this paper, we study two approaches to overcome the missing values problem in dissimilarity-based classification of multi-way data. Namely, imputation by factorization, and a modification of the previously proposed Continuous Multi-way Shape measure for comparing multi-way objects.

In this paper we define distance functions for data sets in a reproduncing kernel Hilbert space (RKHS) context. To this aim we introduce kernels for data sets that provide a metrization of the power set. The proposed distances take into account the underlying generating probability distributions. In particular, we propose kernel distances that rely on the estimation of density level sets of the underlying data distributions, and that can be extended from data sets to probability measures. The performance of the proposed distances is tested on several simulated and real data sets.

Bi-clustering, or co-clustering, refers to the task of finding sub-matrices (indexed by a group of columns and a group of rows) within a matrix such that the elements of each sub-matrix are related in some way, for example, that they are similar under some metric. As in traditional clustering, a crucial parameter in bi-clustering methods is the number of groups that one expects to find in the data, something which is not always available or easy to guess. The present paper proposes a novel method for performing bi-clustering based on the concept of low-rank sparse non-negative matrix factorization (S-NMF), with the additional benefit that the optimum rank

k

is chosen automatically using a minimum description length (MDL) selection procedure, which favors models which can represent the data with fewer bits. This MDL procedure is tested in combination with three different S-NMF algorithms, two of which are novel, on a simulated example in order to assess the validity of the procedure.

This paper introduces a novel spectral clustering approach based on kernels to analyze time-varying data. Our approach is developed within a multiple kernel learning framework, which, in this case is assumed as a linear combination model. To perform such linear combination, weighting factors are estimated by a ranking procedure yielding a vector calculated from the eigenvectors-derived-clustering-method. Particularly, the method named kernel spectral clustering is considered. Proposed method is compared to some conventional spectral clustering techniques, namely, kernel k-means and min-cuts. Standard k-means as well. The clustering performance is quantified by the normalized mutual information and Adjusted Rand Index measures. Experimental results prove that proposed approach is an useful tool for both tracking and clustering dynamic data, being able to manage applications for human motion analysis.

Feature selection is an essential preprocessing step for classifiers with high dimensional training sets. In pattern recognition, feature selection improves the performance of classification by reducing the feature space but preserving the classification capabilities of the original feature space. Image classification using frequent approximate subgraph mining (FASM) is an example where the benefits of features selections are needed. This is due using frequent approximate subgraphs (FAS) leads to high dimensional representations. In this paper, we explore the use of feature selection algorithms in order to reduce the representation of an image collection represented through FASs. In our results we report a dimensionality reduction of over 50% of the original features and we get similar classification results than those reported by using all the features.

Learning in datasets that suffer from imbalanced class distribution is an important problem in Pattern Recognition. This paper introduces a novel algorithm for data balancing, based on compact set clustering of the majority class. The proposed algorithm is able to deal with mixed, as well as incomplete data, and with arbitrarily dissimilarity functions. Numerical experiments over repository databases show the high quality performance of the method proposed in this paper according to area under the ROC curve and imbalance ratio.

Instance selection methods get low accuracy in problems with imbalanced databases. In the literature, the problem of imbalanced databases has been tackled applying oversampling or undersampling methods. Therefore, in this paper, we present an empirical study about the use of oversampling and undersampling methods to improve the accuracy of instance selection methods on imbalanced databases. We apply different oversampling and undersampling methods jointly with instance selectors over several public imbalanced databases. Our experimental results show that using oversampling and undersampling methods significantly improves the accuracy for the minority class.

Fuzzy Cognitive Maps (FCM) are a proper knowledge-based tool for modeling and simulation. They are denoted as directed weighted graphs with feedback allowing causal reasoning. According to the transformation function used for updating the activation value of concepts, FCM can be grouped in two large clusters: discrete and continuous. It is notable that FCM having discrete outputs never exhibit chaotic states, but this premise can not be ensured for FCM having continuous output. This paper proposes a learning methodology based on Swarm Intelligence for estimating the most adequate transformation function for each map neuron (concept). As a result, we can obtain FCM showing better stability properties, allowing better consistency in the hidden patterns codified by the map. The performance of the proposed methodology is studied by using six challenging FCM concerning the field of the HIV protein modeling.

Data mining for the discovery of novel, useful patterns, encounters obstacles when dealing with high-dimensional datasets, which have been documented as the “curse” of dimensionality. A strategy to deal with this issue is the decomposition of the input feature set to build a multi-classifier system. Standalone decomposition methods are rare and generally based on random selection. We propose a decomposition method which uses information theory tools to arrange input features into uncorrelated and relevant subsets. Experimental results show how this approach significantly outperforms three baseline decomposition methods, in terms of classification accuracy.

Feature Selection in very-high-dimensional or small sample problems is particularly prone to computational and robustness complications. It is common to resort to feature ranking approaches only or to randomization techniques. A recent novel approach to the randomization idea in form of Dependency-Aware Feature Ranking (DAF) has shown great potential in tackling these problems well. Its original definition, however, leaves several technical questions open. In this paper we address one of these questions: how to define stopping rules of the randomized computation that stands at the core of the DAF method. We define stopping rules that are easier to interpret and show that the number of randomly generated probes does not need to be extensive.

The selection of prototypes for the dissimilarity space is a key aspect to overcome problems related to the curse of dimensionality and computational burden. How to properly define and select the prototypes is still an open issue. In this paper, we propose the selection of clusters as prototypes to create low-dimensional spaces. Experimental results show that the proposed approach is useful in the problems presented. Especially, the use of the minimum distances to clusters for representation provides good results.

Attribute reduction is an important issue in classification problems. This paper proposes a novel method for categorizing attributes in a decision table based on transforming the binary discernibility matrix into a simpler one called basic binary discernibility matrix. The effectiveness of the method is theoretically demonstrated. Experiments show application results of the proposed method.

Contrast pattern miners and contrast pattern classifiers typically use a quality measure to evaluate the discriminative power of a pattern. Since many quality measures exist, it is important to perform comparative studies among them. Nevertheless, previous studies mostly compare measures based on how they impact the classification accuracy. In this paper, we introduce a comparative study of quality measures over different aspects: accuracy using the whole training set, accuracy using pattern subsets, and accuracy and compression for filtering patterns. Experiments over 10 quality measures in 25 repository databases show that there is a huge correlation among different quality measures and that the most accurate quality measures are not appropriate in contexts like pattern filtering.

A common problem with feature selection is to establish how many features should be retained at least so that important information is not lost. We describe a method for choosing this number that makes use of Support Vector Machines. The method is based on controlling an angle by which the decision hyperplane is tilt due to feature selection.

Experiments were performed on three text datasets generated from a Wikipedia dump. Amount of retained information was estimated by classification accuracy. Even though the method is parametric, we show that, as opposed to other methods, once its parameter is chosen it can be applied to a number of similar problems (e.g. one value can be used for various datasets originating from Wikipedia). For a constant value of the parameter, dimensionality was reduced by from 78% to 90%, depending on the data set. Relative accuracy drop due to feature removal was less than 0.5% in those experiments.

Designing reliable computer-aided diagnosis (CADx) systems based on data extracted from breast images and patient data to provide a second opinion to radiologists is still a challenging and yet unsolved problem. This paper proposes two benchmarking datasets (one of them representative of low resolution digitized Film Mammography images and the other one representative of high resolution Full Field Digital Mammography images) aimed to (1) modeling and exploring machine learning classifiers (MLC); (2) evaluating the impact of mammography image resolution on MLC; and (3) comparing the performance of breast cancer CADx methods. Also, we include a comparative study of four groups of image-based descriptors (intensity, texture, multi-scale texture and spatial distribution of the gradient), and combine them with patient’s clinical data to classify masses. Finally, we demonstrate that this combination of clinical data and image descriptors is advantageous in most CADx scenarios.

Multi-label learning has been becoming an increasingly active area into the machine learning community since a wide variety of real world problems are naturally multi-labeled. However, it is not uncommon to find disparities among the number of samples of each class, which constitutes an additional challenge for the learning algorithm.

Smote

is an oversampling technique that has been successfully applied for balancing single-labeled data sets, but has not been used in multi-label frameworks so far. In this work, several strategies are proposed and compared in order to generate synthetic samples for balancing data sets in the training of multi-label algorithms. Results show that a correct selection of seed samples for oversampling improves the classification performance of multi-label algorithms. The uniform generation oversampling, provides an efficient methodology for a wide scope of real world problems.

The paper presents an online matrix factorization algorithm for multilabel learning. This method addresses the multi-label annotation problem finding a joint embedding that represents both instances and labels in a common latent space. An important characteristic of the novel method is its scalability, which is a consequence of its formulation as an online learning algorithm. The method was systematically evaluated in different standard datasets and compared against state-of-the-art space embedding multi-label learning algorithms showing competitive results.

Although the general relevance of Testor Theory as the theoretical ground for useful feature selection procedures is well known, there are no practical means, nor any standard methodologies, for assessing the behavior of a testor-finding algorithm when faced with specific circumstances. In this work, we present a practical framework, with proven theoretical foundation, for assessing the behavior of both deterministic and meta-heuristic testor-finding algorithms when faced with specific phenomena.

The purpose of CBIR (Content Based Image Retrieval) systems is to allow users to retrieve pictures related to a semantic concept of their interest, when no other information but the images themselves is available. Commonly, a series of images are presented to the user, who judges on their relevance. Several different models have been proposed to help the construction of interactive systems based on relevance feedback. Some of these models consider that an optimal query point exists, and focus on adapting the similarity measure and moving the query point so that it appears close to the relevant results and far from those which are non-relevant. This implies a strong causality between the low level features and the semantic content of the images, an assumption which does not hold true in most cases. In this paper, we propose a novel method that considers the search as a multi-objective optimization problem. Each objective consists of minimizing the distance to one of the images the user has considered relevant. Representatives of the Pareto set are considered as points of interest in the search space, and parallel searches are performed for each point of interest. Results are then combined and presented to the user. A comparatively good performance has been obtained when evaluated against other baseline methods.

This paper presents a novel method to address the problem of indexing a large set of images taking advantage of associated multimodal content such as text or tags. The method finds relationships between the visual and text modalities enriching the image content representation to improve the performance of content-based image search.

This method finds a mapping that connects visual and text information that allows to project new (annotated and unannotated) images to the space defined by semantic annotations, this new representation can be used to search into the collection using a query-by-example strategy and to annotate new unannotated images. The principal advantage of the proposed method is its formulation as an online learning algorithm, which can scale to deal with large image collections. The experimental evaluation shows that the proposed method, in comparison with several baseline methods, is faster and consumes less memory, keeping a competitive performance in content-based image search.

Queries dealing with complex data, such as images, face semantic problems that might compromise results quality. Such problems have their source on the differences found between the semantic interpretation of the data and their low level machine code representation. The descriptors utilized in such representation translate intrinsic characteristics of the data (usually color, shape and texture) into qualifying attributes. Different descriptors represent different intrinsic characteristics that can get different aspects of the data while processing a similarity comparison among them. Therefore, the use of multiple descriptors tends to improve data separation and categorization, if compared to the use of a single descriptor. Another relevant fact is that some specific intrinsic characteristics are essential for identifying a subset of the data. Based on such premises, this work proposes the use of boundary conditions to identify image subsets and then use the best descriptor combination for each of these subsets aimed at decreasing the existing “semantic gap”. Throughout the conducted experiments, the use of the proposed technique had better results when compared to individual descriptor use (employing the same boundary conditions) and to various descriptors combination without the use of boundary conditions.

Image segmentation is a critical step in computer vision tasks constituting an essential issue for pattern recognition and visual interpretation. In this paper, we propose a new stopping criterion for the mean shift iterative algorithm by using images defined in ℤ

n

ring, with the goal of reaching a better segmentation. We carried out also a study on the weak and strong of equivalence classes between two images. An analysis on the convergence with this new stopping criterion is carried out too.

The impact of using evolutionary optimised wavelet subband stuctures as allowed in JPEG2000 Part 2 in polar iris image compression is investigated. The recognition performance of two different feature extraction schemes applied to correspondingly compressed images is compared to the usage of the dyadic decomposition structure of JPEG2000 Part 1 in the compression stage. Recognition performance is significantly improved, provided that the image set used in evolutionary optimisation and actual application is identical. Generalisation to different settings (individuals, sample acquisition conditions, feature extraction techniques) is found to be low.

Image Segmentation and Automatic Image Annotation are two research fields usually addressed independently. Treating these problems simultaneously and taking advantage of each other’s information may improve their individual results. In this work our ultimate goal is image annotation, which we perform using the hierarchical structure of irregular pyramids. We propose a new criterion to create new segmentation levels in the pyramid using low-level cues and semantic information coming from the annotation step. Later, we use the improved segmentation to obtain better annotation results in an iterative way across the hierarchy.We perform experiments in a subset of the Corel dataset, showing the relevance of combining both processes to improve the results of the final annotation.

In this paper, we review some alternatives to reduce the computational complexity of the Non-Local Means image filter and present a CUDA-based implementation of it for GPUs, comparing its performance on different GPUs and with respect to reference CPU implementations. Starting from a naive CUDA implementation, we describe different aspects of CUDA and the algorithm itself that can be leveraged to decrease the execution time. Our GPU implementation achieved speedups of up to 35.8x with respect to our reduced-complexity reference implementation on the CPU, and more than 700x over a plain CPU implementation.

In this paper we propose a new method to automatically correct wide-angle lens distortion from the distorted lines generated by the projection on the image of 3D straight lines. We have to deal with two major problems: on the one hand, wide-angle lenses produce a strong distortion, which makes the detection of distorted lines a particularly difficult task. On the other hand, the usual single parameter polynomial lens distortion models is not able to manage such a strong distortion. We propose an extension of the Hough transform by adding a distortion parameter to detect the distorted lines, and division lens distortion models to manage wide-angle lens distortion. We present some experiments on synthetic and real images to show the ability of the proposed approach to automatically correct this type of distortion. A comparison with a state-of-the-art method is also included to show the benefits of our method.

Survey. The main goal of the Algebraic Approach is the design of a unified scheme for the representation of objects for the purposes of their recognition and the transformation of such representations in the suitable algebraic structures. It makes possible to develop corresponding regular structures ready for analysis by algebraic, geometrical and topological techniques. Development of this line of image analysis and pattern recognition is of crucial importance for automated image mining and application problems solving. It is selected and briefly characterized main aspects of current state of the image analysis algebraization. Special attention is paid to the recent results of the Russian mathematical school.

Nowadays a huge volume of data (e.g. images and videos) are daily generated in several areas. The importance of this subject has led to a new paradigm known as eScience. In this scenario, the biological image domain emerges as an important research area given the great impact that it can leads in real solutions and people’s lives. On the other hand, to cope with this massive data it is necessary to integrate into the same environment not only several techniques involving image processing, description and classification, but also feature selection methods. Hence, in the present paper we propose a new framework capable to join these techniques in a single and efficient pipeline, in order to characterize biological images. Experiments, performed with the ImageCLEF dataset, have shown that the proposed framework presented notable results, reaching up to 87.5% of accuracy regarding the plant species classification, which is highly relevant and a non-trivial task.

Various vision applications exploit matching algorithms to locate a target object in a scene image. A new fast matching algorithm based on recursive calculation of oriented gradient histograms over several circular sliding windows is presented. In order to speed up the algorithm pyramidal image decomposition technique and parallel implementation with modern multicore processors are utilized. The proposed fast algorithm yields a good invariance performance for both in-plane and out-of-plane rotations of a scene image. Computer results obtained with the proposed algorithm are presented and compared with those of common algorithms in terms of matching accuracy and processing time.

In document analysis line segmentation is a necessary prerequisite step for further analysing of textual components. While much work has been devoted to line segmentation of regular text documents, this work can not be easily adopted to documents that contain specialist components such as tables or mathematical expressions. In this paper we concentrate on a line segmentation technique for documents containing mathematical expressions, which, due to their two dimensional structure are often comprised of multiple distinct lines. We present an approach to line segmentation in the presence of mathematics that is based on a set of histogram measures and heuristics considering vertical and horizontal distances of characters only. The method also provides a technique to distinguish consecutive lines that are vertically overlapped but belong to different mathematical expressions. Experiments on data sets of 200 and 1000 maths pages, respectively, show a high rate of accuracy.

Object detection is an important and challenging task in computer vision. In cascaded detectors, a scanned image is passed through a cascade in which all stage detectors have to classify a found object positively. Common detection algorithms use a sliding window approach, resulting in multiple detections of an object. Thus, the merging of multiple detections is a crucial step in post-processing which has a high impact on the final detection performance. First, this paper proposes a novel method for merging multiple detections that exploits intra-cascade confidences using Dempster’s Theory of Evidence. The evidence theory allows hereby to model confidence and uncertainty information to compute the overall confidence measure for a detection. Second, this confidence measure is applied to improve the accuracy of the determined object position. The proposed method is evaluated on public object detection benchmarks and is shown to improve the detection performance.

Image exams are a fundamental tool in health care for decision making. A challenge in Content-based image retrieval (CBIR) is to provide a timely answer that complies with the specialist’s expectation. There are different systems with different techniques to CBIR in literature. However, even with so much research, there are still particular challenges to be overcame, such as the semantic gap. In this paper, we presented a new spinal-image comparison method based on the perception of specialists during his/her analysis of spine lesions. We take advantage of a color extractor and propose a shape descriptor considering the visual patterns that the radiologists use to recognize anomalies in images. The experiments shown that our approach achieved promising results, testifying that the automatic comparison of images should consider all relevant visual aspects and comparisons’ criteria, which are defined by the specialists.

We proposed a new segmentation based on Active Contours (AC) for vector-valued image that incorporates evidential distance. The proposed method combine both Belief Functions (BFs) and probability functions in the Bhattacharyya distance framework. This formulation allows all features issued from vector-valued image and guide the evolution of AC using an inside/outside descriptor. The imprecision caused by the variation of the contrast issued from the multiple channels is incorporated in the BFs as weighted parameters. We demonstrated the performance of the proposed algorithm using some challenging color biomedical images.

Monitoring wildlife populations is important to assess ecosystem health, attend environmental protection activities and undertake research studies about ecology. However, the traditional techniques are temporally and spatially limited; in order to extract information quickly and accurately about the current state of the environment, processing and recognition of acoustic signals are used. In the literature, several research studies about automatic classification of species through their vocalizations are found; however, in many of them the segmentation carried out in the preprocessing stage is briefly mentioned and, therefore, it is difficult to be reproduced by other researchers. This paper is specifically focused on detection of regions of interest in the audio recordings. A methodology for threshold estimation in segmentation techniques based on energy of a frequency band of a birdsong recording is described. Experiments were carried out using chunks taken from the RMBL-Robin database; results showed that a good performance of segmentation can be obtained by computing a threshold as a linear function where the independent variable is the estimated noise.

This paper refers to the application of higher-order statistical signal processing techniques (cumulant calculation) on Gaussian noise cancellation. The performed procedure, joined to a convolution process and Fast Fourier Transform (FFT) application, results in the complete estimation (i.e., amplitude, frequency and phase recovery) of any corrupted periodic signal. Whereas tone frequency estimation is performed by 4th-order cumulant calculation, phase recovery is achieved by the convolution of the cumulant calculation and the corrupted signal. At last, the original signal amplitude is recovered by means of modification of the resulting amplitude spectrum. In this paper, higher-order statistics foundations are presented and the validation of the proposed algorithm is revealed in both theoretical and practical sense. Obtained results are highly satisfactory.

In this work, we explore the integration of hierarchical Language Models (HLMs) in different modules of a Spoken Dialog System. First of all, HLMs are integrated into the Automatic Speech Recognition system. In order to carry out this integration, within the recognition process, finite-state machines were considered. This approach was compared to a two step decoding process in which HLMs are used to rescore a graph. Then, HLMs were also used for Language Understanding (LU) purposes. Two architectures were compared theoretically and empirically in both ASR and LU modules.

Query-by-Example Spoken Term Detection (QbE-STD) tasks are usually addressed by representing speech signals as a sequence of feature vectors by means of a parametrization step, and then using a pattern matching technique to find the candidate detections. In this paper, we propose a phoneme-based approach in which the acoustic frames are first converted into vectors representing the

a posteriori

probabilities for every phoneme. This strategy is specially useful when the language of the task is a priori known. Then, we show how this representation can be used for QbE-STD using both a Segmental Dynamic Time Warping algorithm and a graph-based method. The proposed approach has been evaluated with a QbE-STD task in Spanish, and the results show that it can be an adequate strategy for tackling this kind of problems.

Artifacts should be corrected previous heart rate variability analysis. A new method for artifact correction in multilead ECG is proposed in this paper. The method detects artifacts in the RR series, takes the corresponding segment of the multilead ECG, uses entropy of the signal for selecting the “cleanest” ECG channel, and uses the wavelet transform to recalculate positions of R peaks. The method was evaluated with ECG records of arrhythmia database MIT/BIH, with good results.

The presence of an unusual high level of turbulent noise in voice signals is related to air leakage in the glottis as a result of incomplete closure of the vocal cords. Harmonics to Noise Ratio (HNR) is an acoustic measure that intends to appraise the amount of that turbulent noise. Several algorithms have been proposed in both time and frequency domain to estimate HNR. The Generalized Variogram (GV) is a time-domain technique proposed for HNR estimation based on a similitude function between two speech windows. The drawbacks of the GV are related to the biased estimation of the amplitude ratio and the final HNR value. The present work deals with these limitations and proposes unbiased estimators. The experimental results show that the described improvements outperform the original GV.

In this paper, we propose a robust environmental sound spectrogram classification approach; its purpose is surveillance and security applications based on the reassignment method and log-Gabor filters. Besides, the reassignment method is applied to the spectrogram to improve the readability of the time-frequency representation, and to assure a better localization of the signal components. In this approach the reassigned spectrogram is passed through a bank of 12 log-Gabor filter concatenation applied to three spectrogram patches, and the outputs are averaged and underwent an optimal feature selection procedure based on a mutual information criterion. The proposed method is tested on a large database consists of 1000 environmental sounds belonging to ten classes. The averaged recognition accuracy is of order 90.87% which obtained using the multiclass support vector machines (SVM’s).

Electrocardiographic (ECG) signals are emerging as a recent trend in the field of biometrics. In this paper, we propose a novel ECG biometric system that combines clustering and classification methodologies. Our approach is based on dominant-set clustering, and provides a framework for outlier removal and template selection. It enhances the typical workflows, by making them better suited to new ECG acquisition paradigms that use fingers or hand palms, which lead to signals with lower signal to noise ratio, and more prone to noise artifacts. Preliminary results show the potential of the approach, helping to further validate the highly usable setups and ECG signals as a complementary biometric modality.

Traditional methodologies use electrocardiographic (ECG) signals to develop automatic methods for onset and peak detection on the arterial pulse wave. In the present work a Multilayer Perceptron (MLP) neural network is used for classifying fiducial points on photoplethysmographic (PPG) signals. System was trained with a dataset of temporal segments from signals located based on information about onset and peak points. Different segments sizes and units in the neural network were used for the classification, and optimal values were searched. Results of the classification reach 98.1% in worse of cases. This proposal takes advantages from MLP neural networks for pattern classification. Additionally, the use of ECG signal was avoided in the presented methodology, making the system robust, less expensive and portable in front of this problem.

Most common approaches to phonotactic language recognition deal with phone decoders as tokenizers. However, units that are not linked to phonetic definitions can be more universals, and therefore conceptually easier to adopt. It is assumed that the overall sound characteristics of all spoken languages can be covered by a broad collection of acoustic units, which can be characterized by acoustic segments. In this paper, such acoustic units, highly desirables for a more general language characterization, are delimited and clustered using Gaussian Mixture Model. A new segmentation method on acoustic units of the speech is proposed for later Gaussian modelling, looking for substitute the phonetic recognizer. This tokenizer is trained over untranscribed data, and it precedes the statistical language modeling phase.