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

Object classification and detection are two fundamental problems in computer vision and pattern recognition. In this paper, we discuss these two research topics, including their backgrounds, challenges, recent progress and our solutions which achieve excellent performance in PASCAL VOC competitions on object classification and detection. Moreover, potential directions are outlined for future research.

There is an increasing demand for a new measure of convexity for discrete sets for various applications. For example, the well-known measures for h-, v-, and hv-convexity of discrete sets in binary tomography pose rigorous criteria to be satisfied. Currently, there is no commonly accepted, unified view on what type of discrete sets should be considered nearly hv-convex, or to what extent a given discrete set can be considered convex, in case it does not satisfy the strict conditions. We propose a novel directional convexity measure for discrete sets based on various properties of the configuration of 0s and 1s in the set. It can be supported by proper theory, is easy to compute, and according to our experiments, it behaves intuitively. We expect it to become a useful alternative to other convexity measures in situations where the classical definitions cannot be used.

Uterine Cervical Cancer is one of the most common forms of cancer in women worldwide. Papanicolau smear test is a well-known screening method of detecting abnormalities in the uterine cervix cells. In this paper we address the problem of anomaly detection in pap smear images. Our method avoids modeling the normal pap smear images which is a very complex task due to the large within class variance of the normal target appearance patterns. The problem is posed as a Visual Attention Mechanism. Indeed the human vision system actively seeks interesting regions in images to reduce the search export in tasks, such as anomaly detection. In this paper, we develop a new method for identifying salient regions in pap smear images and compare this to two previously reported approaches. We then consider how such machine-saliency methods can be used to improve human performance in a realistic anomaly detection task.

Noninvasive imaging of unstained living cells allows to study living specimens without altering them, and is a widely used technique in biotechnology for determining biological and biochemical roles of proteins. Fluorescence and contrast images are both used complementarily for better outcomes. However, segmentation of contrast images is particularly difficult due to the presence of lighting/shade-off artifacts, defocused scans, or overlapping. In this work, we make use of the optical properties intervening during the image formation process for cell segmentation. We propose the shear oriented polar snakes, an active contour model that implicitly involves phase information. Experimental results confirms the method suitability for cell images segmentation.

The problem of detecting chemical activity in drugs from its molecular description constitutes a challenging and hard learning task. The corresponding prediction problem can be tackled either as a binary classification problem (active versus inactive compounds) or as a one class problem. The first option leads usually to better prediction results when measured over small and fixed databases while the second could potentially lead to a much better characterization of the active class which could be more important in more realistic settings. In this paper, a comparison of these two options is presented when support vector models are used as predictors.

Urdu Language is written in Nastalique writing style, which is highly cursive, context sensitive and is difficult to process as only the last character in its ligature resides on the baseline. This paper focuses on the development of OCR using Hidden Markov Model and rule based post-processor. The recognizer gets the main body (without diacritics) as input and recognizes the corresponding ligature. Accuracy of the system is 92.73% for printed and then scanned document images at 36 font size.

Induction motors are the most common engine used worldwide. When they are summited to extensive working journals, e.g. in industry, faults may appear, generating a performance reduction on them. Several works have been focused on detecting early mechanical and electrical faults before damage appears in the motor. However, the main drawback of them is the complexity on the motor’s signal mathematical processing. In this paper, a new methodology is proposed for detecting misalignment faults in induction motors. Through signal vibration and orbital analysis, misalignment faults are studied, generating characteristically patterns that are used for fault identification. Artificial Neural Networks are evolved with an evolutionary algorithm for misalignment pattern recognition, using two databases (training and recovering respectively). The results obtained, indicate a good performance of Artificial Neural Networks with low confusion rates, using experimental patterns obtained from real situations where motors present a certain level of misalignment.

In this work we explore the use of computer vision for bus detection in the context of intelligent transport systems. We propose a simple and efficient method to detect moving objects using a probabolistic modelling of the scene. For classification of the detected moving regions we study the use of eigenfaces.

This paper presents a novel approach for automatic music genre recognition in the visual domain that uses two texture descriptors. For this, the audio signal is converted into spectrograms and then textural features are extracted from this visual representation. Gabor filters and LPQ texture descriptors were used to capture the spectrogram content. In order to evaluate the performance of local feature extraction, some different zoning mechanisms were taken into account. The experiments were performed on the Latin Music Database. At the end, we have shown that the SVM classifier trained with LPQ is able to achieve a recognition rate above 80%. This rate is among the best results ever presented in the literature.

We assess the feasibility of unseen appliance recognition through the analysis of their electrical signatures recorded using low-cost smart plugs. By unseen, we stress that our approach focuses on the identification of appliances that are of different brands or models than the one in training phase. We follow a strictly defined protocol in order to provide comparable results to the scientific community. We first evaluate the drop of performance when going from seen to unseen appliances. We then analyze the results of different machine learning algorithms, as the k-Nearest Neighbor (k-NN) and Gaussian Mixture Models (GMMs). Several tunings allow us to achieve 74% correct accuracy using GMMs which is our current best system.

Wind power generation is a green solution to power generation that is receiving increasing interest worldwide. Wind speed forecasting is critical for this technology to succeed and remains today as a challenge to the research community. This paper presents a neural network fusion approach to multi-step-ahead, short-term forecasting of wind speed time-series. Wind speed forecasts are generated using a bank of neural networks that combine predictions from three different forecasters. The wind speed forecasters include a naïve model; a physical model and a custom designed artificial neural network model. Data used in the experiments are telemetric measurements of weather variables from wind farms in Eastern Canada, covering the period from November 2011 to October 2012. Our results show that the combination of three different forecasters leads to substantial performance improvements over recommended reference models.

The green coverage region is a relevant information to be extracted from remote sensing agriculture images. Automatic methods based on threshold and vegetation indices are often applied to address this task. However, sub-orbital remote sensing images have elements that can hinder the automatic analysis. Also, supervised methods can suffer from imbalance since there is often many more green coverage samples available than regions of gaps, weed and degraded areas. We propose an anomaly detection approach to deal with these challenges. Parametric anomaly detection methods using the normal distribution were used and compared with vegetation indices, unsupervised and supervised learning methods. The results showed that anomaly detection algorithms can handle better the green coverage detection. The proposed methods showed similar or better accuracy when compared with the competing methods. It deals well with different images and with the imbalance problem, confirming the practical application of the approach.

Physically impaired people may use Surface Electromyography (sEMG) signals to control assistive devices in an automatic way. sEMG signals directly reflect the human motion intention, they can be used as input information for active exoskeleton control. This paper proposes a set of myoelectric algorithms based on machine learning for detecting movement intention aimed at controlling an upper limb active exoskeleton. The algorithms use a feature extraction stage based on a combination of time and frequency domain features (mean absolute value – waveform length, and auto-regressive model, respectively). The pattern recognition stage uses Linear Discriminant Analysis, K-Nearest Neighbor, Support Vector Machine and Bayesian classifiers. Additionally, two post-processing techniques are incorporated: majority vote and transition removal. The performance of the algorithms is evaluated with parameters of sensitivity, specificity, positive predictive value, error rate and active error rate, under typical conditions. These evaluations allow identifying pattern recognition algorithms for real-time control of an active exoskeleton.

Most of the state-of-the-art Arabic Optical Character Recognition systems use Hidden Markov Models to model Arabic characters. Much of the attention is paid to provide the HMM system with new features, pre-processing, or post-processing modules to improve the performances. In this paper, we present an Arabic OCR system using Restricted Boltzmann Machines (RBMs) to model Arabic characters. The recently announced ALTEC dataset for typewritten OCR system is used to train and test the system. The results show a 26% increase in the average word accuracy rate and 8% increase in the average character accuracy rate compared to the HMM system.

In this work we present a system for autonomous discovery of asteroids, space trash and other moving objects. This system performs astronomical image data reduction based on an image processing pipeline. The processing steps of the pipeline include astrometric and photometric reduction, sequence alignment, moving object detection and astronomical analysis, making the system capable of discovering and monitoring previously unknown moving objects in the night sky.

While a number of techniques have been developed for table recognition in ordinary text documents, very little work has been done on tables that contain mathematical expressions. The latter problem is complicated by the fact that mathematical formulae often have a tabular layout themselves, thus not only blurring the distinction between table and content structure, but often leading to a number of possible, equally valid interpretations. However, a reliable understanding of the layout of a formula is often a necessary prerequisite to further semantic interpretation. In this paper, a graph representation for complex mathematical table structures is presented. A set of rewriting rules is applied to the graph allows for reliable re-composition of cells in order to identify several valid table interpretations. The effectiveness of the technique is demonstrated by applying it to a set of mathematical tables from standard text books that has been manually ground-truthed.

The topology of complex brain networks allows efficient dynamic interactions between spatially distinct regions. Neuroimaging studies have provided consistent evidence of dysfunctional connectivity among the cortical circuitry in Parkinson’s disease; however, little is known about the topological properties of brain networks underlying these alterations. This paper introduces a methodology to explore aberrant changes in hierarchical patterns of nodal centrality through cortical networks, combining graph theoretical analysis and morphometric connectivity. The edges in graph were estimated by correlation analysis and thresholding between 148 nodes defined by cortical regions. Our findings demonstrated that the networks organization was disrupted in the patients with PD. We found a reconfiguration in hierarchical weighting of high degree hubs in structural networks associated with levels of cognitive decline, probably related to a system-wide compensatory mechanism. Simulated targeted attack on the network’s nodes as measures of network resilience showed greater effects on information flow in advanced stages of disease.

Currently, material resistance research is looking for biomaterials where mechanical properties (like fatigure resistance) and biocompatibility are the main characteristics to take into account. To understand the behavior of materials subject to fatigue, usually we analyze how the material responds to cyclic forces. Failures due to fatigue are the first cause of cracks in materials. Normally, failures start with a superficial deficiency and produce micro cracks, which grow until a total break of the material. In this work we deal with the early detection of micro cracks on the surface of bone cement, while they are under fatigue tests, in order to characterize the material and design better and more resistant materials according to where they would be applied. The method presented for crack detection consists in several stages: noise reduction, shadow elimination, image segmentation and path detection for crack analysis. At the end of the analysis of one image, the number of cracks and the length of each one can be obtained (based on the maximum length of crack candidates). If a video is analyzed, the evolution of cracks in the material can be observed.

Mild cognitive impairment (MCI) is common in Parkinson’s Disease (PD) patients and it is key to predict the development of dementia. There is not report of discriminant accuracy for MCI using based-surface cortical morphometry. This study used Cortical-Thickness (CT) combined to Local-Gyrification-Index (LGI) to assess discriminant accuracy for MCI stages in PD. Sixty-four patients with idiopathic PD and nineteen healthy controls (HC) were analyzed. CT and LGI were estimated using Freesurfer software. Principal Component Analysis and Lineal Discriminant Analysis (LDA) assuming a common diagonal covariance matrix (or Naive-Bayes classifier) was used with cross-validation leave-one-subject-out scheme. Accuracy, sensibility and specificity were reported to different classification analysis. CT combined to LGI limited revealed the best discrimination with accuracy of 82,98%, sensitivity of 85.71% and specificity of 80.77%. A validation process using independent and more heterogeneous data set and further longitudinal studies, are necessary to confirm our results.

Training systems based on virtual reality are used in several areas of human activities. In some kinds of training is important to know the trainee’s skills. It can be done in those systems but requires high-end computers to achieve good performance. Recently, the use of embedded systems connected to the training system was proposed for training assessment, with the goal of decreasing requirements of the main system. However, some questions are still open and a deep study of this proposal was not performed. This paper provides answers for some of those questions.

This work proposes an improvement of the multi-objective evolutionary method for the protein residue-residue contact prediction called MECoMaP. This method bases its prediction on physico-chemical properties of amino acids, structural features and evolutionary information of the proteins. The evolutionary algorithm produces a set of decision rules that identifies contacts between amino acids. These decision rules generated by the algorithm represent a set of conditions to predict residue-residue contacts. A new encoding used, a fast evaluation of the examples from the training data set and a treatment of unbalanced classes of data were considered to improve the the efficiency of the algorithm.

Histology images are used to identify biological structures present in living organisms — cells, tissues, and organs — correctly. The structure of tissues varies according to the type and purpose of the tissue. Automatic identification of tissues is an open problem in image processing. In this paper, the identification of loose connective and muscle tissues based on morphological tissue information is presented.

Image identification is commonly evaluated in isolation. This is done either by eye or via some other quality measure. Expert criteria — by eye — are used to evaluate the identification results. Experimental results show that the proposed approach yields results close to the real results, according to expert opinion.

Computer tomographic colonography, combined with computer-aided detection, is a promising emerging technique for colonic polyp analysis. We present a complete pipeline for polyp detection, starting with a simple colon segmentation technique that enhances polyps, followed by an adaptive-scale candidate polyp delineation and classification based on new texture and geometric features that consider both the information in the candidate polyp and its immediate surrounding area. The proposed system is tested with ground truth data, including challenging flat and small polyps. For polyps larger than 6

mm

in size we achieve 100% sensitivity with just 0.9 false positives per case, and for polyps larger than 3

mm

in size we achieve 93% sensitivity with 2.8 false positives per case.

Several antiviral drugs have been approved for treating HIV infected patients. These drugs inhibit the function of proteins which are essential in the virus life cycle, thus preventing the virus reproduction. However, due to its high mutation rate the HIV is capable to develop resistance to administered therapy. For this reason, it is important to study the resistance mechanisms of the HIV proteins in order to make a better use of existing drugs and design new ones. In the last ten years, numerous statistical and machine learning approaches were applied for predicting drug resistance from protein genome information. In this paper we first review the most relevant techniques reported for addressing this problem. Afterward, we describe a Fuzzy Cognitive Map based modeling which allows representing the causal interactions among the protein positions and their influence on the resistance. Finally, an extended comparison experimentation is carried out, which reveals that this model is competitive with well-known approaches and notably outperforms other techniques from literature.

After pre-processing and segmenting suspicious masses in mammographies based on the Top-Hat and Markov Random Fields methods, we developed a mass-detection algorithm that uses gray level co-occurrence matrices, gray level difference statistics, gray level run length statistics, shape descriptors and intensity parameters as the entry of a vector support machine classifier. During the classification process we test up to 63 image features, keeping the 35 most important and obtaining 85% of accuracy score.

This paper presents a novel method that follows the right ventricle (RV) shape during a whole cardiac cycle in magnetic resonance sequences (MRC). The proposed approach obtains an initial coarse segmentation by a bidirectional per pixel motion descriptor. Then a refined segmentation is obtained by fusing the previous segmentation with geometrical observations at each frame. A main advantage of the proposed approach is a robust MRI heart characterization without any prior information. The proposed approach achieves a Dice Score of 0.62 evaluated over 32 patients.

In recent years, with the advent of High-resolution Computed Tomography (HRCT), there has been an increased interest for diagnosing Chronic Obstructive Pulmonary Disease (COPD), which is commonly presented as emphysema. Since low-attenuation areas in HRCT images describe different emphysema patterns, the discrimination problem should focus on the characterization of both local intensities and global spatial variations. We propose a novel texture-based classification framework using complex Gabor filters and local binary patterns. We also analyzed a set of global and local texture descriptors to characterize emphysema morphology. The results have shown the effectiveness of our proposal and that the combination of descriptors provides robust features that lead to an improvement in the classification rate.

The Papanicolaou test is used for early prediction of cervical cancer. Computer vision techniques for automating the microscopy analysis of cervical cells in this test have received great attention. Cell segmentation is needed here in order to obtain appropriate features for classification of abnormal cells. However, accurate segmentation of the cell cytoplasm is difficult, due to cell overlapping and variability of color and intensity. This has determined a growing interest in classifying cells using only features from the nuclei, which are easier to segment. In this work, we classified cells in the pap-smear test using a combination of morphometric and Haralick texture features, obtained from the nucleus gray-level co-occurrence matrix. A comparison was made among various classifiers using these features and data dimensionality reduction through PCA. The results obtained showed that this combination can be a promising alternative in order to automate the analysis of cervical cells.

This study focuses on the effects of artificial cardiac tissue in the excitation-contraction process of the ventricular muscle. We developed a spatio-temporal computerized model of the whole heart that handles half millimeter sized compartments using 1 microsecond time step. We employed the effect of muscle fiber direction, laminar sheets, depolarization period and other parameters. The artificial tissue differs from the normal one in several ways, so their describing parameters are also modified. In our simulation the depolarization wave (DW) conduction speed of the artificial tissue was decreased by up to 3 times. In presence of a two centimeter wide and 2 mm thick artificial tissue slice, the maximal depolarization delay was 38 msec. Large ventricle size, low conducting speed and spaciousness of the injured ventricular tissue are the main generating factors of arrhythmia, while the location of the artificial tissue has secondary importance.

Quantifying concentrations of target molecules near cellular structures, within cells or tissues, requires identifying the gold particles in immunogold labelled images. In this paper, we address the problem of automatically detect them accurately and reliably across multiple scales and in noisy conditions. For this purpose, we introduce a new contrast filter, based on an adaptive version of the H-extrema algorithm. The filtered images are simplified with a geodesic reconstruction to precisely segment the candidates. Once the images are segmented, we extract classical features and then classify using the majority vote of multiple classifiers. We characterize our algorithm on a pilot data and present results that demonstrate its effectiveness.

The goal of this research is to use a bionic model to enhance classifi- cation of Dysarthria. The model based on the main features of the mammalian olfactory system is the initial stage of the recognition process. The bionic mod- el aimed to achieve an enhancement in the separation ability of the dysarthric features. The recognition performance obtained by four different pattern recog- nition algorithms using the bionic model to improve the features is shown and discussed. The results indicated that bionic model had clear influence on classi- fication performance of well-known techniques using dysarthria database as case study. We regard the results of this study as a promising initial step to the use of bionic model as a recognition improvement function.

Applying heart rate variability (HRV) analysis on ambulatory ECG monitoring is a very useful decision support tool for cardiovascular diagnosis. The presence of non-valid beats (artefacts) on the RR interval time-series affects the diagnosis accuracy using this technique. Despite the importance of artefacts recognition prior to exclusion, no paper was found characterizing quantitatively the performance of, on the one hand, the extracted features and, on the other hand, the clustering methods on artefacts recognition for HRV analysis. In this paper we evaluate the performance of several combinations of three feature extraction methods and four clustering methods (based on machine learning techniques) for the artefacts beats recognition on the ECG signal. The trade-off between performance indexes suggests the use of a non-linear principal component analysis as feature extraction method and a multilayer perceptron (MLP) as clustering method, with sensitivity, specificity and positive-predictive-value (PPV) equal to respectively 95 %, 95.9 % and 98 %.

In this paper a binary biometric comparator based on Counting Bloom filters is introduced. Within the proposed scheme binary biometric feature vectors are analyzed and appropriate bit sequences are mapped to Counting Bloom filters. The comparison of resulting sets of Counting Bloom filters significantly improves the biometric performance of the underlying system. The proposed approach is applied to binary iris-biometric feature vectors, i.e. iris-codes, generated from different feature extractors. Experimental evaluations, which carried out on the CASIA-v3-Interval iris database, confirm the soundness of the presented comparator.

In this paper, we focus on gender recognition in challenging large scale scenarios. Firstly, we review the literature results achieved for the problem in large datasets, and select the currently hardest dataset: The Images of Groups. Secondly, we study the extraction of features from the face and its local context to improve the recognition accuracy. Different descriptors, resolutions and classifiers are studied, overcoming previous literature results, reaching an accuracy of 89.8%.

This paper focuses on the analysis of the i-vector paradigm, a compact representation of spoken utterances that is used by most of the state of the art speaker verification systems. This work was mainly motivated by the need to quantify the impact of their steps on the final performance, especially their ability to model data according to a theoretical Gaussian framework. These investigations allow to highlight the key points of the approach, in particular a core conditioning procedure, that lead to the success of the i-vector paradigm.

A new method for action recognition is proposed by revisiting LBP-based dynamic texture operators. It captures the similarity of motion around keypoints tracked by a realtime semi-dense point tracking method. The use of self-similarity operator allows to highlight the geometric shape of rigid parts of foreground object in a video sequence. Inheriting from the efficient representation of LBP-based methods and the appearance invariance of patch matching method, the method is well designed for capturing action primitives in unconstrained videos. Action recognition experiments, made on several academic action datasets validate the interest of our approach.

Palmprint identification is still considered as a challenging research line in Biometrics. Nowadays, the performance of this techniques highly depends on the quality of the involved palmprints, specially if the identification is performed in latent palmprints. In this paper, we propose a new feature model for representing palmprints and dealing with the problems of missing and spurious minutiae. Moreover, we propose a novel verification algorithm based in this feature model, which uses a strategy for finding adaptable local matches between substructures obtained from images. In experimentation, we show that our proposal achieves highest scores in latent palmprint matching, improving some of the best results reported in the literature.

A critical element in multi-biometrics systems, is the rule to fuse the information from the different sources. The component sub-systems are often designed to further produce indices of input image quality and/or of system reliability. These indices can be used as weights assigned to scores (weighted fusion) or as a selection criterion to identify the subset of systems that actually take part in a single fusion operation. Many solutions rely on the estimation of the joint distributions of conditional probabilities of the scores from the single subsystems. The negative counterpart is that such very effective solutions require training and a high number of training samples, and also assume that score distributions are stable over time. In this paper we propose a unified representation of the score and of the quality/reliability index that simplifies the process of fusion, provides performance comparable to those currently offered by top performing schemes, yet does not require a prior estimation of score distributions. This is an interesting feature in highly dynamic systems, where the set of relevant subjects may undergo significant variations across time.

While combining more than one biometric sample, recognition algorithm, modality or sensor, commonly referred to as multi-biometrics, is common practice to improve accuracy of biometric systems, fusion at segmentation level has so far been neglected in literature. This paper introduces the concept of multi-segmentation fusion for combining independent iris segmentation results. Fusion at segmentation level is useful to (1) obtain more robust recognition rates compared to single segmentation; (2) avoid additional storage requirements compared to feature-level fusion, and (3) save processing time compared to employing parallel chains of feature-extractor dependent segmentation. As proof of concept, manually labeled segmentation results are combined using the proposed technique and shown to increase recognition accuracy for representative algorithms on the well-known CASIA-V4-Interval dataset.

Gesture recognition has become a very active research area with the advent of the Kinect sensor. The most common approaches for gesture recognition use temporal information and are based on methods such as Hidden Markov Models (HMM) and Dynamic Time Warping (DTW). In this paper, we present a novel non-temporal alternative for gesture recognition using the Microsoft Kinect device. The proposed approach, Recognition by Characteristic Window (RCW), identifies, using clustering techniques and a sliding window, distinctive portions of individual gestures which have low overlapping information with other gestures. Once a distinctive portion has been identified for each gesture, all these sub-sequences are used to recognize a new instance. The proposed method was compared against HMM and DTW on a benchmark gesture’s dataset showing very competitive performance.

The automatic identification of kinship relations from pairs of facial images is an emerging research area in pattern analysis with possible applications in image retrieval and annotation, forensics and historical studies. This work explores the computer identification of pairs of kins using different facial features, based on geometric and textural data, and state-of-the-art classifiers. We first analyzed different facial attributes individually, selecting the most effective feature variables with a two stage feature selection algorithm. Then, these features were combined together, selecting again the most relevant ones. Experiments shows that the proposed approach provides a valuable solution to the kinship verification problem, as suggested by the comparison with a different method on the same data and on the same experimental protocol.

Biometric recognition is still a very difficult task in real-world scenarios wherein unforeseen changes in degradations factors like noise, occlusion, blurriness and illumination can drastically affect the extracted features from the biometric signals. Very recently Haar-like rectangular features which have usually been used for object detection were introduced for biometric recognition resulting in systems that are robust against most of the mentioned degradations [9]. The problem with these features is that one can define many different such features for a given biometric signal and it is not clear whether all of these features are required for the actual recognition or not. This is exactly what we are dealing with in this paper: How can an initial set of Haar-like rectangular features, that have been used for biometric recognition, be reduced to a set of most influential features? This paper proposes total sensitivity analysis about the mean for this purpose for two different biometric traits, iris and face. Experimental results on multiple public databases show the superiority of the proposed system, using the found influential features, compared to state-of-the-art biometric recognition systems.

This paper presents a method for classification of imbalanced splice-site classification problems, the proposed method consists of the generation of artificial instances that are incorporated to the dataset. Additionally, the method uses a genetic algorithm to introduce just instances that improve the performance. Experimental results show that the proposed algorithm obtains a better accuracy to detect splice-sites than other implementations on skewed data-sets.

The applications of Support Vector Machines (SVM) in speaker recognition are mainly related to Gaussian Mixtures and Universal Background Model based supervector paradigm. Recently, has been proposed a new approach that allows represent each acoustic frame in a binary discriminant space. Also a representation of a speaker - called accumulative vectors - obtained from the binary space has been proposed. In this article we show results obtained using SVM with the accumulative vectors and Nuisance Attribute Projection (NAP) as a method for compensating the session variability. We also introduce a new method to counteract the effects of the signal length in the conformation of the accumulative vectors to improve the performance of SVM.

Fusion of several biometric traits has traditionally been regarded as more secure than unimodal recognition systems. However, recent research works have proven that this is not always the case. In the present article we analyse the performance and robustness of several fusion schemes to indirect attacks. Experiments are carried out on a multimodal system based on face and iris, a user-friendly trait combination, over the publicly available multimodal Biosecure DB. The tested system proves to have a high vulnerability to the attack regardless of the fusion rule considered. However, the experiments prove that not necessarily the best fusion rule in terms of performance is the most robust to the type of attack considered.

In this paper, a topological approach for gait-based gender recognition is presented. First, a stack of human silhouettes, extracted by background subtraction and thresholding, were glued through their gravity centers, forming a 3D digital image

I

. Second, different

filters

(i.e. particular orders of the simplices) are applied on ∂ 

K

(

I

) (a simplicial complex obtained from

I

) which capture relations among the parts of the human body when walking. Finally, a

topological signature

is extracted from the persistence diagram according to each filter. The measure cosine is used to give a similarity value between topological signatures. The novelty of the paper is a notion of robustness of the provided method (which is also valid for gait recognition). Three experiments are performed using all human-camera view angles provided in CASIA-B database. The first one evaluates the named topological signature obtaining 98.3% (lateral view) of correct classification rates, for gender identification. The second one shows results for different human-camera distances according to training and test (i.e. training with a human-camera distance and test with a different one). The third one shows that upper body is more discriminative than lower body.

With the introduction of template protection techniques, privacy and security of biometric data have been enforced. Meeting the required properties of irreversibility, i.e. avoiding a reconstruction of original biometric features, and unlinkability among each other, template protection can enhance security of existing biometric systems in case tokens are stolen. However, with increasing resolution and number of enrolled users in biometric systems, means to compress biometric signals become an imminent need and practice, raising questions about the impact of image compression on recognition accuracy of template protection schemes, which are particularly sensitive to any sort of signal degradation. This paper addresses the important topic of iris-biometric fuzzy commitment schemes’ robustness with respect to compression noise. Experiments using a fuzzy commitment scheme indicate, that medium compression does not drastically effect key retrieval performance.

Due to the large areas covered by surveillance systems, employed cameras usually lack intersection of field of view, refraining us from mapping the location of a person in a camera to another one. Therefore, when a subject appears in a camera, a person re-identification method is required to discover whether the subject has been previously identified in a different camera. Even though several approaches have been proposed in the literature, person re-identification is still a challenging problem due to appearance variation between cameras, changes in illumination, pose variation, and low quality data, among others. To reduce the effect of the aforementioned difficulties, we propose a person re-identification approach that models the appearance of the subjects based on multiple samples collected from multiple cameras and employs person detection and tracking to enhance the robustness of the method. Experiments conducted on three public available data sets demonstrate improvements over existing methods.

This paper proposes the introduction of annular Zernike polynomials for representing iris images data. This representation offers notables advantages like representing the images on a continuous domain that allows the application of Functional Data Analysis techniques, preserving their original nature. In addition, it provides a significant dimensionality reduction of the data, while it still has a high discriminative power. The proposed approach also deals with the occlusion problems that can be present in this type of images. In order to corroborate the effectiveness of the introduced approach, identification experiments were carried out. Iris international databases were used. Some of them are characterized by the presence of severe occlusion problems. Results have shown high recognition accuracy.

This paper reports an analysis of the benefits of using color information on a region-based face recognition system. Three different color spaces are analysed (

RGB

,

YC

b

C

r

,

lαβ

) in a very challenging scenario matching good quality mugshot images against video surveillance images. This scenario is of special interest for forensics, where examiners carry out a comparison of two face images using the global information of the faces, but paying special attention to each individual facial region (eyes, nose, mouth, etc.). This work analyses the discriminative power of 15 facial regions comparing both the grayscale and color information. Results show a significant improvement of performance when fusing several regions of the face compared to just using the whole face image. A further improvement of performance is achieved when color information is considered.

The accurate localization of facial features is an important task for the face recognition process. One of the most used approaches to achieve this goal is the Active Shape Models (ASM) method and its different extensions. In this work, a new method is proposed for obtaining a Local Binary Patterns (LBP) based profile for representing the local appearance of landmark points of the shape model in ASM. The experimental evaluation, conducted on XM2VTS and BioID databases, shows the good performance of the proposal.

This paper proposes an approach for detecting important parts of faces in uncontrolled imaging settings. Regions of special interest in faces of humans are eyes and eyebrows, nose and mouth. The approach works by first extracting ORB (Oriented FAST and Rotated BRIEF) and SURF (Speeded up robust features) features, secondly a supervised learning step with a random subset of images is performed using k-means algorithm for devising the clusters’ centers of the important parts of faces. For the testing set of images the normalized values of each new ORB or SURF feature is weighted positively depending on its similarity and proximity of a cluster center (a face part). Tests were performed using the BioID dataset which consists of 1521 images of 23 different subjects in a variety of situations. Results show that the use of ORB features for face parts localization is more efficient and more precise than SIFT or SURF features alone. Also, the relative spatial weighting of a combination of ORB and SURF features enhances the localization of parts of faces.

We present a novel multi-shot re-identification method, that merges together two different pattern recognition paradigms for describing objects: feature-based and relation-based. The former aims at encoding visual properties that characterize the object per se. The latter gives a relational description of the object considering how the visual properties are interdependent. The method considers SDALF as feature-based description: SDALF segregates salient body parts, exploiting symmetry and asymmetry principles. Afterwards, the parts are described by color, texture and region-based features. As relation-based description we consider the covariance of features, recently employed for re-identification: in practice, the parts found by SDALF are additionally encoded as covariance matrices, capturing structural properties otherwise missed. The resulting descriptor, dubbed SDALF+C, is superior to SDALF by about 2% and to the covariance-based description by a 53%, both in terms of average rank1 probability, considering 5 different multi-shot benchmark datasets (i-LIDS, ETHZ1,2,3 and CAVIAR4REID).

Segmentation of image sequences is a challenging task in computer vision. Time-of-Flight cameras provide additional information, namely depth, that can be integrated as an additional feature in a segmentation approach. Typically, the depth information is less sensitive to environment changes. Combined with appearance, this yields a more robust segmentation method. Motivated by the fact that a simple combination of two information sources might not be the best solution, we propose a novel scheme based on Dempster’s theory of evidence. In contrast to existing methods, the use of Dempster’s theory of evidence allows to model inaccuracy and uncertainty. The inaccuracy of the information is influenced by an adaptive weight, that provides a measurement of how reliable a certain information might be. We compare our method with others on a publicly available set of image sequences. We show that the use of our proposed fusion scheme improves the segmentation.

Early gesture recognition consists of recognizing gestures at their beginning, using incomplete information. Among other applications, these methods can be used to compensate for the delay of gesture-based interactive systems. We propose a new approach for early recognition of full-body gestures based on dynamic time warping (DTW) that uses a single example from each category. Our method is based on the comparison between time sequences obtained from known and unknown gestures. The classifier provides a response before the unknown gesture finishes. We performed experiments in the MSR-Actions3D benchmark and another data set we built. Results show that, in average, the classifier is capable of recognizing gestures with 60% of the information, losing only 7.29% of accuracy with respect to using all of the information.

Video-based Augmented Reality (VAR) aims to add 3D virtual objects (3D VOs) to a real world video sequence, in order to provide additional and useful information to facilitate some tasks, like computer aided surgery, simulation in a real environment, satellite positioning, interior design, among others. To achieve a consistent and convincing augmented scene, it is necessary that the VOs are properly occluded by real objects (Occlusion Problem in VAR); in this paper, we present a strategy based on the use of the

Kinect

sensor to solve this problem. In the occlusion stage we evaluate distances between real and VOs. Then, the parts of the VO occluded by a real object are calculated and removed. We found that the

Kinect

sensor is appropriate to be used for handling occlusions in indoor environments, dynamic scenarios and real-time applications. Experiments showed comparable results with the state of the art in both issues: occlusion handling and processing time.

A reliable system for recognition and tracking of a moving target in nonuniformly illuminated scenes is presented. The system employs a filter bank of space-variant correlation filters adapted to local statistical parameters of the observed scene in each frame. When a scene frame is captured, a fragment of interest is constructed in the frame around predicted location of the target based on a kinematic model. The fragment is firstly pointwise processed to correct the illumination. Afterwards, the state of the target is estimated from the restored fragment by employing a bank of space-variant correlation filters. The performance of the proposed system in terms of object recognition and tracking is tested with nonuniformly illuminated and noisy scenes. The results are compared with those of common techniques based on correlation filtering.

In this work we propose an algorithm for film restoration aimed at reducing the flicker effect while preserving the original overall illumination of the film. We also present a comparative study of the performance of this algorithm implemented following a sequential approach on a CPU and following a parallel approach on a GPU using OpenCL.

Most of the action recognition methods presented in the literature cannot be applied to real life situations. Some of them demand expensive feature extraction or classification processes, some require previous knowledge about starting and ending action times, others are just not scalable. In this paper, we present a real time action recognition method that uses information about the variation of the silhouette shape, which can be extracted and processed with little computational effort, and we apply a fast configuration of lightweight classifiers. The experiments are conducted on theWeizmann dataset and show that our method achieves the state-of-the-art accuracy in real time and can be scaled to work on different conditions and be applied several times simultaneously.

The constant increase in the availability of digital videos has demanded the development of techniques capable of managing these data in a faster and more efficient way, especially concerning the content analysis. One of the research areas that have recently evolved significantly at this point is video summarization, which consists of generating a short version of a certain video, such that the users can grasp the central message transmitted by the original video. Many of the video summarization approaches make use of clustering algorithms, with the goal of extracting the most important frames of the videos to compose the final summary. However, special clustering algorithms based on a spectral approach have obtained superior results than those obtained with classical clustering algorithms, not only in video summarization techniques but also in other fields, such as machine learning, pattern recognition, and data mining. This work proposes a method for summarization of videos, regardless of their genre, using spectral clustering algorithms. Possibilities of algorithm parallelization for the purpose of optimizing the general performance of the proposed methodology are also discussed.

We present an approach for motion estimation from videos captured by depth-sensing cameras. Our method uses the technique of graph matching to find groups of pixels that move to the same direction in subsequent frames. In order to choose the best matching for each patch, we minimize a cost function that accounts for distances on RGB and XYZ spaces. Our application runs at real-time rates for low resolution images and has shown to be a convenient framework to deal with input data generated by the new depth-sensing devices. The results show clearly the advantage obtained in the use of RGB-D images over RGB images.

A methodology for automatic segmentation and classification of multi-channel data related to motion capture (MoCap) videos of cyclic activities are presented. Regarding this, a kernel approach is employed to obtain a time representation, which captures the cyclic behavior of a given multi-channel data. Moreover, we calculate a mapping based on kernel principal component analysis, in order to obtain a low-dimensional space that encodes the main cyclic behaviors. From such, low-dimensional space the main segments of the studied activity are inferred. Then, a distance based classifier is used to classified each MoCap video segment. A well-known MoCap database is tested which contains different activities performed by humans. Attained results shows how our approach is a simple alternative to obtain a suitable classification performance in comparison to complex methods for MoCap analysis.

Graph models offer high representational power and useful structural cues. Unfortunately, tracking objects by matching graphs over time is in general NP-hard. Simple appearance-based trackers are able to find temporal correspondences fast and efficient, but often fail to overcome challenging situations like occlusions, distractors and noise. This paper proposes an approach, where an attributed graph is used to represent the structure of the target object and multiple, simple trackers in combination with structural cues replace the costly graph matching. Thus, the strengths of both methodologies are combined to overcome their weaknesses. Experiments based on synthetic videos are used to evaluate two possible structural cues. Results show the superiority of the cue based on barycentric coordinates and the potential of the proposed tracking approach in challenging situations.

In video-conference and distance learning videos, the moment that someone makes a hand-raising gesture is relevant to be included in the video annotation. However, gesture recognition can be challenging in such scenarios. We propose a system to detect faces, the hand-raising gesture and annotate the video. The Lienhart-Maydt object detection method is used, in which each frame is classified. Then, the gesture is detected by analyzing intervals of frames. Our approach was tested in videos with several characteristics. The results show that our method can deal with illumination and background variations, is able to detect multiple gestures and it is robust to confusing gestures. Besides it allow the use of moving cameras.

We show that

the way

people observe video sequences, other than

what

they observe, is important for the understanding and the prediction of human activities. In this study, we consider 36 surveillance videos, organized in four categories (

confront, nothing, fight, play

): the videos are observed by 19 people, ten of them are experienced operators and the other nine are novices, and the gaze trajectories of both populations are recorded by an eye tracking device. Due to the proved superior ability of experienced operators in predicting violence in surveillance footage, our aim is to distinguish the two classes of people, highlighting in which respect expert operators differ from novices. Extracting spatio-temporal features from the eye tracking data, and training standard machine learning classifiers, we are able to discriminate the two groups of subjects with an average accuracy of 80.26%. The idea is that expert operators are more focused on few regions of the scene, sampling them with high frequency and low predictability. This can be thought as a first step toward the advanced automated analysis of video surveillance footage, where machines imitate as best as possible the attentive mechanisms of humans.

In the last years, the learning from multi-label data has attracted significant attention from a lot of researchers, motivated from an increasing number of modern applications that contain this type of data. Several methods have been proposed for solving this problem, however how to make feature weighting on multi-label data is still lacking in the literature. In multi-label data, each data point can be attributed to multiple labels simultaneously, thus a major difficulty lies in the determinations of the features useful for all multi-label concepts. In this paper, a new method for feature weighting in multi-label learning area is presented, based on the principles of the well-known ReliefF algorithm. The experimental stage shows the effectiveness of the proposal.

This paper presents an annotation tool that detects entities in the biomedical domain. By enriching the lexica of the Freeling analyzer with bio-medical terms extracted from dictionaries and ontologies as SNOMED CT, the system is able to automatically detect medical terms in texts. An evaluation has been performed against a manually tagged corpus focusing on entities referring to pharmaceutical drug-names, substances and diseases. The obtained results show that a good annotation tool would help to leverage subsequent processes as data mining or pattern recognition tasks in the biomedical domain.

In this work we propose the intensive use of embedded memory blocks and logic blocks of the FPGA device for signature matching. In our approach we arrange signatures in memory arrays (MA) of embedded memory blocks, so that every signature is matched in one clock cycle. The matching logic is shared among all the signatures in one MA. In addition, we propose a character recodification method that allows memory bits savings, leading to a low byte/character cost. For fast memory addressing we employ the unique substring detection, in doing so we process four bytes per clock cycle while hardware replication is significantly reduced.

We cast the problem of discovering the community structure in networks as the composition of community candidates, obtained from several community detection base algorithms, into a coherent structure. In turn, this composition can be cast into a maximum-weight clique problem, and we propose an ant colony optimization algorithm to solve it. Our results show that the proposed method is able to discover better community structures, according to several evaluation criteria, than the ones obtained with the base algorithms. It also outperforms, both in quality and in speed, the recently introduced FG-Tiling algorithm.

Computing the minimum spanning tree (MST) is a common task in the pattern recognition and the computer vision fields. However, little work has been done on efficient general methods for solving the problem on large datasets where graphs are complete and edge weights are given implicitly by a distance between vertex attributes. In this work we propose a generic algorithm that extends the classical Boruvka’s algorithm by using nearest neighbors search structures to significantly reduce time and memory consumption. The algorithm can also compute in a straightforward way approximate MSTs thus further improving speed. Experiments show that the proposed method outperforms classical algorithms on large low-dimensional datasets by several orders of magnitude.