Showing papers in "Pattern Analysis and Applications in 2019"

Micro-expression recognition based on 3D flow convolutional neural network

Abstract: Micro-expression recognition (MER) is a growing field of research which is currently in its early stage of development. Unlike conventional macro-expressions, micro-expressions occur at a very short duration and are elicited in a spontaneous manner from emotional stimuli. While existing methods for solving MER are largely non-deep-learning-based methods, deep convolutional neural network (CNN) has shown to work very well on such as face recognition, facial expression recognition, and action recognition. In this article, we propose applying the 3D flow-based CNNs model for video-based micro-expression recognition, which extracts deeply learned features that are able to characterize fine motion flow arising from minute facial movements. Results from comprehensive experiments on three benchmark datasets—SMIC, CASME/CASME II, showed a marked improvement over state-of-the-art methods, hence proving the effectiveness of our fairly easy CNN model as the deep learning benchmark for facial MER.

An implementation of optimized framework for action classification using multilayers neural network on selected fused features

Abstract: In video sequences, human action recognition is a challenging problem due to motion variation, in frame person difference, and setting of video recording in the field of computer vision. Since last few years, applications of human activity recognition have increased significantly. In the literature, many techniques are implemented for human action recognition, but still they face problem in contrast of foreground region, segmentation, feature extraction, and feature selection. This article contributes a novel human action recognition method by embedding the proposed frames fusion working on the principle of pixels similarity. An improved hybrid feature extraction increases the recognition rate and allows efficient classification in the complex environment. The design consists of four phases, (a) enhancement of video frames (b) threshold-based background subtraction and construction of saliency map (c) feature extraction and selection (d) neural network (NN) for human action classification. Results have been tested using five benchmark datasets including Weizmann, KTH, UIUC, Muhavi, and WVU and obtaining recognition rate 97.2, 99.8, 99.4, 99.9, and 99.9%, respectively. Contingency table and graphical curves support our claims. Comparison with existent techniques identifies the recognition rate and trueness of our proposed method.

Feature selection via Lèvy Antlion optimization

Abstract: In this paper, a modification of the newly proposed antlion optimization (ALO) is introduced and applied to feature selection relied on the Levy flights. ALO method is one of the encouraging swarm intelligence algorithms which make use of random walking to perform the exploration and exploitation operations. Random walks based on uniform distribution is responsible for premature convergence and stagnation. A Levy flight random walk is suggested as a permutation for performing a local search. Levy random walking grants the optimization ability to generate several solutions that are apart from existing solutions and furthermore enables it to escape from local minima and much efficient in examining large search area. The proposed Levy antlion optimization (LALO) algorithm is applied in a wrapper-based mode to select optimal feature combination that maximizing classification accuracy while minimizing the number of selected features. LALO algorithm is applied on 21 different benchmark datasets against genetic algorithm (GA), particle swarm optimization (PSO), and the native ALO methods. Different initialization methods and several evaluation criteria are employed to assess algorithm diversification and intensification of the optimization algorithms. The experimental results demonstrate the significant improvement in the proposed LALO over the native ALO and many well-known methods used in feature selection.

A review of retinal blood vessels extraction techniques: challenges, taxonomy, and future trends

Abstract: The visual exploration of retinal blood vessels assists ophthalmologists in the diagnoses of different abnormalities of the eyes such as diabetic retinopathy, glaucoma, cardiovascular ailment, high blood pressure, arteriosclerosis, and age-related macular degeneration. The manual inspection of retinal vasculature is an extremely challenging and tedious task for medical experts due to the complex structure of an eye, tiny blood vessels, and variation in vessels width. Several automatic retinal vessels extraction techniques have been proposed in contemporary literature, which assist ophthalmologists in the timely identification of an eye disorders. However, due to the fast evolution of such techniques, a comprehensive survey is needed. This survey presents a comprehensive review of such techniques, strategies, and algorithms presented to date. The techniques are classified into logical groups based on the underlying methodology employed for retinal vessel extraction. The performance of existing techniques is reported on the publicly accessible datasets in term of various performance measures, providing a valuable comparison among the techniques. Thus, this survey presents a valuable resource for researchers working toward automatic extraction of retinal vessels.

A comprehensive study of clustering ensemble weighting based on cluster quality and diversity

Abstract: Clustering as a major task in data mining is responsible for discovering hidden patterns in unlabeled datasets. Finding the best clustering is also considered as one of the most challenging problems in data mining. Due to the problem complexity and the weaknesses of primary clustering algorithm, a large part of research has been directed toward ensemble clustering methods. Ensemble clustering aggregates a pool of base clusterings and produces an output clustering that is also named consensus clustering. The consensus clustering is usually better clustering than the output clusterings of the basic clustering algorithms. However, lack of quality in base clusterings makes their consensus clustering weak. In spite of some researches in selection of a subset of high quality base clusterings based on a clustering assessment metric, cluster-level selection has been always ignored. In this paper, a new clustering ensemble framework has been proposed based on cluster-level weighting. The certainty amount that the given ensemble has about a cluster is considered as the reliability of that cluster. The certainty amount that the given ensemble has about a cluster is computed by the accretion amount of that cluster by the ensemble. Then by selecting the best clusters and assigning a weight to each selected cluster based on its reliability, the final ensemble is created. After that, the paper proposes cluster-level weighting co-association matrix instead of traditional co-association matrix. Then, two consensus functions have been introduced and used for production of the consensus partition. The proposed framework completely overshadows the state-of-the-art clustering ensemble methods experimentally.

Training neural networks on high-dimensional data using random projection

Abstract: Training deep neural networks (DNNs) on high-dimensional data with no spatial structure poses a major computational problem. It implies a network architecture with a huge input layer, which greatly increases the number of weights, often making the training infeasible. One solution to this problem is to reduce the dimensionality of the input space to a manageable size, and then train a deep network on a representation with fewer dimensions. Here, we focus on performing the dimensionality reduction step by randomly projecting the input data into a lower-dimensional space. Conceptually, this is equivalent to adding a random projection (RP) layer in front of the network. We study two variants of RP layers: one where the weights are fixed, and one where they are fine-tuned during network training. We evaluate the performance of DNNs with input layers constructed using several recently proposed RP schemes. These include: Gaussian, Achlioptas’, Li’s, subsampled randomized Hadamard transform (SRHT) and Count Sketch-based constructions. Our results demonstrate that DNNs with RP layer achieve competitive performance on high-dimensional real-world datasets. In particular, we show that SRHT and Count Sketch-based projections provide the best balance between the projection time and the network performance.

DK-means: a deterministic K-means clustering algorithm for gene expression analysis

Abstract: Clustering has been widely applied in interpreting the underlying patterns in microarray gene expression profiles, and many clustering algorithms have been devised for the same. K-means is one of the popular algorithms for gene data clustering due to its simplicity and computational efficiency. But, K-means algorithm is highly sensitive to the choice of initial cluster centers. Thus, the algorithm easily gets trapped with local optimum if the initial centers are chosen randomly. This paper proposes a deterministic initialization algorithm for K-means (DK-means) by exploring a set of probable centers through a constrained bi-partitioning approach. The proposed algorithm is compared with classical K-means with random initialization and improved K-means variants such as K-means++ and MinMax algorithms. It is also compared with three deterministic initialization methods. Experimental analysis on gene expression datasets demonstrates that DK-means achieves improved results in terms of faster and stable convergence, and better cluster quality as compared to other algorithms.

Optimized feature selection-based clustering approach for computer-aided detection of lung nodules in different modalities

Abstract: Early detection of pulmonary lung nodules plays a significant role in the diagnosis of lung cancer. Computed tomography (CT) and chest radiographs (CRs) are currently being used by radiologists to detect such nodules. In this paper, we present a novel cluster-based classifier architecture for lung nodule computer-aided detection systems in both modalities. We propose a novel optimized method of feature selection for both cluster and classifier components. For CRs, we make use of an independent database comprising of 160 cases with a total of 173 nodules for training purposes. Testing is implemented on a publicly available database created by the Standard Digital Image Database Project Team of the Scientific Committee of the Japanese Society of Radiological Technology (JRST). The JRST database comprises 154 CRs containing one radiologist-confirmed nodule in each. In this research, we exclude 14 cases from the JRST database that contain lung nodules in the retrocardiac and subdiaphragmatic regions of the lung. For CT scans, the analysis is based on threefold cross-validation performance on 107 cases from publicly available dataset created by Lung Image Database Consortium comprised of 280 nodules. Overall, with a specificity of 3 false positives per case/patient on average, we show a classifier performance boost of 7.7% for CRs and 5.0% for CT scans when compared to a single aggregate classifier architecture.

3D object recognition and classification: a systematic literature review

Abstract: In this paper, we present a systematic literature review concerning 3D object recognition and classification. We cover articles published between 2006 and 2016 available in three scientific databases (ScienceDirect, IEEE Xplore and ACM), using the methodology for systematic review proposed by Kitchenham. Based on this methodology, we used tags and exclusion criteria to select papers about the topic under study. After the works selection, we applied a categorization process aiming to group similar object representation types, analyzing the steps applied for object recognition, the tests and evaluation performed and the databases used. Lastly, we compressed all the obtained information in a general overview and presented future prospects for the area.

A generalized multi-scale line-detection method to boost retinal vessel segmentation sensitivity

Abstract: Many chronic eye diseases can be conveniently investigated by observing structural changes in retinal blood vessel diameters. However, detecting changes in an accurate manner in face of interfering pathologies is a challenging task. The task is generally performed through an automatic computerized process. The literature shows that powerful methods have already been proposed to identify vessels in retinal images. Though a significant progress has been achieved toward methods to separate blood vessels from the uneven background, the methods still lack the necessary sensitivity to segment fine vessels. Recently, a multi-scale line-detector method proved its worth in segmenting thin vessels. This paper presents modifications to boost the sensitivity of this multi-scale line detector. First, a varying window size with line-detector mask is suggested to detect small vessels. Second, external orientations are fed to steer the multi-scale line detectors into alignment with flow directions. Third, optimal weights are suggested for weighted linear combinations of individual line-detector responses. Fourth, instead of using one global threshold, a hysteresis threshold is proposed to find a connected vessel tree. The overall impact of these modifications is a large improvement in noise removal capability of the conventional multi-scale line-detector method while finding more of the thin vessels. The contrast-sensitive steps are validated using a publicly available database and show considerable promise for the suggested strategy.

Boosting sensitivity of a retinal vessel segmentation algorithm

Abstract: The correlation between retinal vessel structural changes and the progression of diseases such as diabetes, hypertension, and cardiovascular problems has been the subject of several large-scale clinical studies. However, detecting structural changes in retinal vessels in a sufficiently fast and accurate manner, in the face of interfering pathologies, is a challenging task. This significantly limits the application of these studies to clinical practice. Though monumental work has already been proposed to extract vessels in retinal images, they mostly lack necessary sensitivity to pick low-contrast vessels. This paper presents a couple of contrast-sensitive measures to boost the sensitivity of existing retinal vessel segmentation algorithms. Firstly, a contrast normalization procedure for the vascular structure is adapted to lift low-contrast vessels to make them at par in comparison with their high-contrast counterparts. The second measure is to apply a scale-normalized detector that captures vessels regardless of their sizes. Thirdly, a flood-filled reconstruction strategy is adopted to get binary output. The process needs initialization with properly located seeds, generated here by another contrast-sensitive detector called isophote curvature. The final sensitivity boosting measure is an adoption process of binary fusion of two entirely different binary outputs due to two different illumination correction mechanism employed in the earlier processing stages. This results in improving the noise removal capability while picking low-contrast vessels. The contrast-sensitive steps are validated on a publicly available database, which shows considerable promise in the strategy adopted in this research work.

A new method for image encryption by 3D chaotic map

Abstract: Image encryption is sensitive to various attacks due to the high volume of the data and similarity of the pixels in different images. Therefore, the encryption algorithms should have high complexity so that analyses become difficult or even impossible. Also it should have lower time order to quickly encrypt high-volume pictures. This paper presents a new method with a simple algorithm, low time order and high level of output complexity. Three nonlinear chaotic sequences which are known as three-dimensional logistic maps are used for image encryption. These three random sequences are extracted and used for pixels rows and columns permutations. Finally, the pixels are changed column by column with third chaotic sequence using XOR operator. This method is compared with well-known algorithms. The results show that the correlation coefficient is improved on average by 0.0028, NPCR and UACI values by 0.09972.

Implications of the curse of dimensionality for supervised learning classifier systems: theoretical and empirical analyses

Abstract: Learning classifier systems are leading evolutionary machine learning systems that employ genetic algorithms to search for a set of optimally general and correct classification rules for a variety of machine learning problems, including supervised classification data mining problems. The curse of dimensionality is a phenomenon that arises when analysing data in high-dimensional spaces. Performance issues when dealing with increasing dimensionality in the training data, such as poor classification accuracy and stalled genetic search, are well known for learning classifier systems. However, a systematic study to establish the relationship between increasing dimensionality and learning challenges in these systems is lacking. The aim of this paper is to analyse the behaviour of Michigan-style learning classifier systems that use the most commonly adopted and expressive interval-based rules representation, under curse of dimensionality (also known as Hughes Phenomenon) problem. In this paper, we use well-established and mathematically founded formal geometrical properties of high-dimensional data spaces and generalisation theory of these systems to propose a formulation of such relationship. The proposed formulations are validated experimentally using a set of synthetic, two-class classification problems. The findings demonstrate that the curse of dimensionality occurs for as few as ten dimensions and negatively affects the evolutionary search with a hyper-rectangular rule representation. A number of approaches to overcome some of the difficulties uncovered by the presented analysis are then discussed. Three approaches are then analysed in more detail using a set of synthetic, two-class classification problems. Experimental study demonstrates the effectiveness of these approaches to handle increasing dimensional data.

Multi-channel local ternary pattern for content-based image retrieval

Abstract: A feature based on a single modality such as color or texture is not sufficient to investigate the appearance variation across multiple images. In this paper, a novel feature referred to as multi-channel local ternary pattern is proposed for image retrieval. The proposed method captures cross-channel color–texture information through the combination of H–V, S–V and V–V channels obtained from HSV representation of the image. Not only texture statistics extracted in this manner contain color information but local texture information is also incorporated in such representations. The effectiveness of the proposed image retrieval method is measured by performing experiments on popular natural, face and texture databases including Corel 1000, Corel 10k, CMU-PIE, STex and MIT VisTex, and results are compared with the existing state-of-the-art techniques. Retrieval results clearly highlight the superior performance of the proposed approach in terms of average precision and average recall.

Bag of recurrence patterns representation for time-series classification

Abstract: Time-series classification (TSC) has attracted a lot of attention in pattern recognition, because wide range of applications from different domains such as finance and health informatics deal with time-series signals. Bag-of-features (BoF) model has achieved a great success in TSC task by summarizing signals according to the frequencies of “feature words” of a data-learned dictionary. This paper proposes embedding the recurrence plots (RP), a visualization technique for analysis of dynamic systems, in the BoF model for TSC. While the traditional BoF approach extracts features from 1D signal segments, this paper uses the RP to transform time-series into 2D texture images and then applies the BoF on them. Image representation of time-series enables us to explore different visual descriptors that are not available for 1D signals and to treat TSC task as a texture recognition problem. Experimental results on the UCI time-series classification archive demonstrates a significant accuracy boost by the proposed bag of recurrence patterns, compared not only to the existing BoF models, but also to the state-of-the art algorithms.

Epileptic seizure detection using fuzzy-rules-based sub-band specific features and layered multi-class SVM

Abstract: In this paper, a new epileptic seizure detection method using fuzzy-rules-based sub-band specific features and layered directed acyclic graph support vector machine (LDAG-SVM) is proposed for classification of electroencephalogram (EEG) signals. Wavelet transformation is used to decompose the input EEG signals into various sub-bands. The nonlinear features, namely approximate entropy, largest Lyapunov exponent and correlation dimension, are extracted from each sub-band. In this proposed work, sub-band specific feature subset that is reduced in size and capable of discriminating samples is selected by employing fuzzy rules. For classification purpose, a new LDAG-SVM is used for detecting epileptic seizure. Every sub-band has its own characteristics. If appropriate features which characterize the specific sub-band are selected, then the classification accuracy is improved and computational complexity is reduced. The important advantage of the fuzzy logic is its close relation to human thinking. Due to the lengthy record and intra-professional variability, automation of epileptic detection is inevitable. Fuzzy rules are the natural choice of employing human expertise to build machine learning system. Performances of the proposed methods are evaluated using two different benchmark EEG datasets, namely Bonn and CHB-MIT. The performance measures such as classification accuracy, sensitivity, specificity, execution time and receiver operating characteristics are used to measure and analyze the performances of the proposed classifier. The proposed LDAG-SVM with fuzzy-rules-based selected sub-band specific features provides better performance in terms of improved classification accuracy with reduced execution time compared to existing methods.

New aggregation operators of single-valued neutrosophic hesitant fuzzy set and their application in multi-attribute decision making

Abstract: The single-valued neutrosophic hesitant fuzzy set (SVNHFS) is a combination of single-valued neutrosophic set and hesitant fuzzy set that is designed for some incomplete, uncertain and inconsistent situations in which each element has a few different values designed by truth membership hesitant function, indeterminacy membership hesitant function and falsity membership hesitant function. In this paper, we define the score function, accuracy function and certainty function of SVNHFS and give the laws to compare the SVNHFS. Then, we propose the single-valued neutrosophic hesitant fuzzy ordered weighted averaging operator and the single-valued neutrosophic hesitant fuzzy hybrid weighted averaging operator and study the properties of the operators. Furthermore, an approach to multi-attribute decision making within the framework of SVNHFS is developed by the proposed aggregation operators. Finally, a practical application of the developed approach is given, and the result shows that our approach is reasonable and effective in dealing with uncertain decision-making problems.

Probabilistic multi-word spotting in handwritten text images

Abstract: Keyword spotting techniques are becoming cost-effective solutions for information retrieval in handwritten documents. We explore the extension of the single-word, line-level probabilistic indexing approach described in our previous works to allow for page-level search of queries consisting in Boolean combinations of several single-keywords. We propose heuristic rules to combine the single-word relevance probabilities into probabilistically consistent confidence scores of the multi-word boolean combinations. An empirical study, also presented in this paper, evaluates the search performance of word-pair queries involving AND and OR Boolean operations. Results of this study support the proposed approach and clearly show its effectiveness. Finally, a web-based demonstration system based on the proposed methods is presented.

Invariant color images representation using accurate quaternion Legendre–Fourier moments

Abstract: Color image representation using quaternion moments has gained more interest during the last few years. In this paper, invariant color image representation using quaternion Legendre–Fourier moments (QLFMs) is presented. A new method for exact, fast and stable computation of QLFMs in polar coordinates is proposed. Explicit rotation, scaling and translation (RST) invariants of quaternion Legendre–Fourier are derived. Numerical experiments are performed to compare the performance of QLFMs with the existing quaternion moments. The comparison clearly shows the superiority of the proposed method over all existing quaternion moments in terms of image reconstruction capability, RST invariances and numerical stability.

Gabor filter-based texture for ancient degraded document image binarization

Abstract: Binarization of ancient degraded document images is a very important step for their preservation and digital use. In this paper, a new simple threshold-based method is proposed for binarization of ancient degraded documents. The proposed method is inspired from the most popular threshold-based methods by exploiting texture information features extracted from both the filtered image using the Gabor filter and the original degraded document. Firstly, a preprocessing stage using the Wiener filter is performed on the degraded image for facilitating the binarization. Then, a Gabor filter bank is weighted according to the dominant slant of the document’s image script for estimating the binarization threshold. Finally, a post-processing stage is applied based on morphological operator for reducing some artifacts. For setting optimal parameters, a new protocol is proposed in the design stage by taking into account the degradation type. Exhaustive experiments are achieved using standard DIBCO datasets series reorganized according to the degradation type and the year of contest. Obtained results are compared against various well-known threshold-based methods. On the other hand, a comparison is achieved with the state-of-the-art methods. Promising results and stability are noticed for the proposed technique, specifically for ink bleed-through degradation and low-contrasted documents.

Sparsity augmented discriminative sparse representation for face recognition

Abstract: Sparse representation-based classification (SRC) has acquired prominent capability in fields of machine learning and pattern recognition. Collaborative representation-based classification (CRC) has achieved a comparable recognition performance with higher speed compared to SRC, which has attracted much attention because it enables to forgo the computationally quite expensive l1-norm sparsity constraint. However, the traditional CRC method neglects the discriminability of representation and recent study has claimed that the sparsity should not be completely neglected for computational costs. In this paper, we propose a sparsity-augmented discriminative sparse representation-based classification method which considers the discriminability and sparsity of representation via augmenting an l2-norm regularization discriminative sparse representation with a computationally inexpensive sparse representation. We utilize an efficient classification method to achieve better performance with a comparable classification time. Experimental results on four face databases show the effectiveness of our proposed method.

Paraphrase plagiarism identification with character-level features

Abstract: Several methods have been proposed for determining plagiarism between pairs of sentences, passages or even full documents. However, the majority of these methods fail to reliably detect paraphrase plagiarism due to the high complexity of the task, even for human beings. Paraphrase plagiarism identification consists in automatically recognizing document fragments that contain reused text, which is intentionally hidden by means of some rewording practices such as semantic equivalences, discursive changes and morphological or lexical substitutions. Our main hypothesis establishes that the original author’s writing style fingerprint prevails in the plagiarized text even when paraphrases occur. Thus, in this paper we propose a novel text representation scheme that gathers both content and style characteristics of texts, represented by means of character-level features. As an additional contribution, we describe the methodology followed for the construction of an appropriate corpus for the task of paraphrase plagiarism identification, which represents a new valuable resource to the NLP community for future research work in this field.

Independent component analysis for a low-channel SSVEP-BCI

Abstract: Generally, the more channels are used to acquire EEG signals, the better the performance of the brain–computer interface (BCI). However, from the user’s point of view, a BCI system comprising a large number of channels is not desirable because of the lower comfort and extended application time. Therefore, the current trend in BCI design is to use the smallest number of channels possible. The problem is, however, that usually when we decrease the number of channels, the interface accuracy also drops significantly. In the paper, we examined whether it is possible to maintain the high accuracy of a BCI based on steady-state visual evoked potentials (SSVEP-BCI) in a low-channel setup using a preprocessing procedure successfully used in a multichannel setting: independent component analysis (ICA). The influence of ICA on the BCI performance was measured in an off-line (24 subjects) mode and online (eight subjects) mode. In the off-line mode, we compared the number of correctly recognized different stimulation frequencies, and in the online mode, we compared the classification accuracy. In both experiments, we noted the predominance of signals that underwent ICA preprocessing. In the off-line mode, we detected 50% more stimulation frequencies after ICA preprocessing than before (in the case of four EEG channels), and in the online mode, we noted a classification accuracy increase of 8%. The most important results, however, were the results obtained for a very low luminance (350 lx), where we noted 71% gain in the off-line mode and 11% gain in the online mode.

Outlier trajectory detection through a context-aware distance

Abstract: This paper presents an original method to detect anomalous human trajectories based on a new and promising context-aware distance. The input of the proposed method is a set of human trajectories from a video surveillance system. A proper representation of each trajectory is developed based on the polar coordinates of the corresponding sub-trajectories. The main focus of the paper is to highlight a context-aware distance between trajectories. This distance implies a weighted average of the differences in the angle, the Euclidean distance, and the number of points in each trajectory. The distance matrix feeds an unsupervised learning method to extract homogeneous groups (clusters) of trajectories. Finally, an outlier detection method is executed over the trajectories in each cluster. The methodology has been empirically evaluated across four experiments with both artificial and real data sets. The tests results have proved promising and illustrate the effectiveness of this approach for anomalous trajectories detection.

Region-based facial representation for real-time Action Units intensity detection across datasets

Abstract: Most research on facial expressions recognition has focused on binary Action Units (AUs) detection, while graded changes in their intensity have rarely been considered. This paper proposes a method for the real-time detection of AUs intensity in terms of the Facial Action Coding System scale. It is grounded on a novel and robust anatomically based facial representation strategy, for which features are registered from a different region of interest depending on the AU considered. Real-time processing is achieved by combining Histogram of Gradients descriptors with linear kernel Support Vector Machines. Following this method, AU intensity detection models are built and validated through the DISFA database, outperforming previous approaches without real-time capabilities. An in-depth evaluation through three different databases (DISFA, BP4D and UNBC Shoulder-Pain) further demonstrates that the proposed method generalizes well across datasets. This study also brings insights about existing public corpora and their impact on AU intensity prediction.

Efficient health-related abnormal behavior detection with visual and inertial sensor integration

Abstract: An increasing number of healthcare issues arise from unsafe abnormal behaviors such as falling and staggering of a rapidly aging population. These abnormal behaviors, often coming with abrupt movements, could potentially be life-threatening if unnoticed; real-time, accurate detection of this sort of behavior is essential for timely response. However, it is challenging to achieve generic, while accurate, abnormal behavior detection in real time with moderate sensing devices and processing power. This paper presents an innovative system as a solution. It utilizes primarily visual data for detecting various types of abnormal behaviors due to accuracy and generality of computer vision technologies. Unfortunately, the volume of the recorded video data is huge, which is preventive to process all in real time. We propose to use elder-carried mobile devices either by a dedicated design or by a smartphone, equipped with inertial sensor to trigger the selection of relevant video data. In this way, the system operates in a trigger verify fashion, which leads to selective utilization of video data to guarantee both accuracy and efficiency in detection. The system is designed and implemented using inexpensive commercial off-the-shelf sensors and smartphones. Experimental evaluations in real-world settings illustrate our system’s promise for real-time accurate detection of abnormal behaviors.

Online streaming feature selection: a minimum redundancy, maximum significance approach

Abstract: All the traditional feature selection methods assume that the entire input feature set is available from the beginning However, online streaming features (OSF) are integral part of many real-world applications In OSF, the number of training examples is fixed while the number of features grows with time as new features stream in A critical challenge for online streaming feature selection (OSFS) is the unavailability of the entire feature set before learning starts OS-NRRSAR-SA is a successful OSFS algorithm that controls the unknown feature space in OSF by means of the rough sets-based significance analysis This paper presents an extension to the OS-NRRSAR-SA algorithm In the proposed extension, the redundant features are filtered out before significance analysis In this regard, a redundancy analysis method based on functional dependency concept is proposed The result is a general OSFS framework containing two major steps, (1) online redundancy analysis that discards redundant features, and (2) online significance analysis, which eliminates non-significant features The proposed algorithm is compared with OS-NRRSAR-SA algorithm, in terms of compactness, running time and classification accuracy during the features streaming The experiments demonstrate that the proposed algorithm achieves better results than OS-NRRSAR-SA algorithm, in every way

Sparse subspace clustering via nonconvex approximation

Abstract: Among existing clustering methods, sparse subspace clustering (SSC) obtains superior clustering performance in grouping data points from a union of subspaces by solving a relaxed $$\ell _{0}$$ -minimization problem by $$\ell _{1}$$ -norm. The use of $$\ell _{1}$$ -norm instead of the $$\ell _{0}$$ one can make the object function convex, while it also causes large errors on large coefficients in some cases. In this work, we propose using the nonconvex approximation to replace $$\ell _{0}$$ -norm for SSC, termed as SSC via nonconvex approximation (SSCNA), and develop a novel clustering algorithm with the enhanced sparsity based on the Alternating Direction Method of Multipliers. We further prove that the proposed nonconvex approximation is closer to $$\ell _{0}$$ -norm than the $$\ell _{1}$$ one and is bounded by $$\ell _{0}$$ -norm. Numerical studies show that the proposed nonconvex approximation helps to improve clustering performance. We also theoretically verify the convergence of the proposed algorithm with a three-variable objective function. Extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed method.

Improving optimum-path forest learning using bag-of-classifiers and confidence measures

Abstract: Machine learning techniques have been actively pursued in the last years, mainly due to the great number of applications that make use of some sort of intelligent mechanism for decision-making processes. In this work, we presented an ensemble of optimum-path forest (OPF) classifiers, which consists into combining different instances that compute a score-based confidence level for each training sample in order to turn the classification process “smarter”, i.e., more reliable. Such confidence level encodes the level of effectiveness of each training sample, and it can be used to avoid ties during the OPF competition process. Experimental results over fifteen benchmarking datasets have shown the effectiveness and efficiency of the proposed approach for classification problems, with more accurate results in more than 67% of the datasets considered in this work. Additionally, we also considered a bagging strategy for comparison purposes, and we showed the proposed approach can lead to considerably better results.

Feature subset selection for classification of malignant and benign breast masses in digital mammography

Abstract: Computer-aided diagnosis of breast cancer is becoming increasingly a necessity given the exponential growth of performed mammograms. In particular, the breast mass diagnosis and classification arouse nowadays a great interest. Texture and shape are the most important criteria for the discrimination between benign and malignant masses. Various features have been proposed in the literature for the characterization of breast masses. The performance of each feature is related to its ability to discriminate masses from different classes. The feature space may include a large number of irrelevant ones which occupy a lot of storage space and decrease the classification accuracy. Therefore, a feature selection phase is usually needed to avoid these problems. The main objective of this paper is to select an optimal subset of features in order to improve masses classification performance. First, a study of various descriptors which are commonly used in the breast cancer field is conducted. Then, selection techniques are used in order to determine the most relevant features. A comparative study between selected features is performed in order to test their ability to discriminate between malignant and benign masses. The database used for experiments is composed of mammograms from the MiniMIAS database. Obtained results show that Gray-Level Run-Length Matrix features provide the best result.