Search

This paper presents a comparison among different texture descriptors and ensembles of descriptors for music genre classification. The features are extracted from the spectrogram calculated starting from the audio signal. The best results are obtained by extracting features from subwindows taken from the entire spectrogram by Mel scale zoning. To assess the performance of our method, two different databases are used: the Latin Music Database (LMD) and the ISMIR 2004 database. The best descriptors proposed in this work greatly outperform previous results using texture descriptors on both databases: we obtain 86.1% accuracy with LMD and 82.9% accuracy with ISMIR 2004. Our descriptors and the MATLAB code for all experiments reported in this paper will be available at https://www.dei.unipd.it/node/2357 .

In the past 10 years virtual reality exposure therapy (VRET) has become a viable alternative for exposure in vivo, the gold standard for the treatment of anxiety disorders. VRET is often regarded as a natural extension of the systematic exposure component of (cognitive) behavior therapy. The objective of this chapter is to discuss the treatment of anxiety disorders with VRET and review research into the outcome of VRET and the presumed (underlying) mechanisms of VRET. We discuss variables, which could play an important role during therapy such as the capability of people to experience the virtual world as real (presence), cognitions, psychophysiology, and the therapeutic alliance. Treatment procedures will be illustrated with case vignettes for different anxiety disorders.

In this paper the problem of finding a face recognition system that works well both under variable illumination conditions and under strictly controlled acquisition conditions is considered. The problem under consideration has to do with the fact that systems that work well (compared with standard methods) with variable illumination conditions often suffer a drop in performance on images where illumination is strictly controlled. In this chapter we review existing techniques for obtaining illumination robustness and propose a method for handling illumination variance that combines different matchers and preprocessing methods. An extensive evaluation of our system is performed on several datasets (CMU, ORL, Extended YALE-B, and BioLab). Our results show that even though some standalone matchers are inconsistent in performance depending on the database, the fusion of different methods performs consistently well across all tested datasets and illumination conditions. Our experiments show that the best result are obtained using gradientfaces as a preprocessing method and orthogonal linear graph embedding as a feature transform.

No abstract.

Convolutional Neural Networks (CNNs) have gained prominence in the research literature on image classification over the last decade. One shortcoming of CNNs, however, is their lack of generalizability and tendency to overfit when presented with small training sets. Augmentation directly confronts this problem by generating new data points providing additional information. In this paper, we investigate the performance of more than ten different sets of data augmentation methods, with two novel approaches proposed here: one based on the Discrete Wavelet Transform and the other on the Constant-Q Gabor transform. Pretrained ResNet50 networks are finetuned on each augmentation method. Combinations of these networks are evaluated and compared across three benchmark data sets of images representing diverse problems and collected by instruments that capture information at different scales: a virus data set, a bark data set, and a LIGO glitches data set. Experiments demonstrate the superiority of this approach. The best ensemble proposed in this work achieves state-of-the-art performance across all three data sets. This result shows that varying data augmentation is a feasible way for building an ensemble of classifiers for image classification (code available at https://github.com/LorisNanni).

In this paper, we examine two strategies for boosting the performance of ensembles of Siamese networks (SNNs) for image classification using two loss functions (Triplet and Binary Cross Entropy) and two methods for building the dissimilarity spaces (FULLY and DEEPER). With FULLY, the distance between a pattern and a prototype is calculated by comparing two images using the fully connected layer of the Siamese network. With DEEPER, each pattern is described using a deeper layer combined with dimensionality reduction. The basic design of the SNNs takes advantage of supervised k-means clustering for building the dissimilarity spaces that train a set of support vector machines, which are then combined by sum rule for a final decision. The robustness and versatility of this approach are demonstrated on several cross-domain image data sets, including a portrait data set, two bioimage and two animal vocalization data sets. Results show that the strategies employed in this work to increase the performance of dissimilarity image classification using SNN are closing the gap with standalone CNNs. Moreover, when our best system is combined with an ensemble of CNNs, the resulting performance is superior to an ensemble of CNNs, demonstrating that our new strategy is extracting additional information.

Traditionally, classifiers are trained to predict patterns within a feature space. The image classification system presented here trains classifiers to predict patterns within a vector space by combining the dissimilarity spaces generated by a large set of Siamese Neural Networks (SNNs). A set of centroids from the patterns in the training data sets is calculated with supervised k-means clustering. The centroids are used to generate the dissimilarity space via the Siamese networks. The vector space descriptors are extracted by projecting patterns onto the similarity spaces, and SVMs classify an image by its dissimilarity vector. The versatility of the proposed approach in image classification is demonstrated by evaluating the system on different types of images across two domains: two medical data sets and two animal audio data sets with vocalizations represented as images (spectrograms). Results show that the proposed system’s performance competes competitively against the best-performing methods in the literature, obtaining state-of-the-art performance on one of the medical data sets, and does so without ad-hoc optimization of the clustering methods on the tested data sets.

In this work, we combine a Siamese neural network and different clustering techniques to generate a dissimilarity space that is then used to train an SVM for automated animal audio classification. The animal audio datasets used are (i) birds and (ii) cat sounds, which are freely available. We exploit different clustering methods to reduce the spectrograms in the dataset to a number of centroids that are used to generate the dissimilarity space through the Siamese network. Once computed, we use the dissimilarity space to generate a vector space representation of each pattern, which is then fed into an support vector machine (SVM) to classify a spectrogram by its dissimilarity vector. Our study shows that the proposed approach based on dissimilarity space performs well on both classification problems without ad-hoc optimization of the clustering methods. Moreover, results show that the fusion of CNN-based approaches applied to the animal audio classification problem works better than the stand-alone CNNs.

Diagnosing pain in neonates is difficult but critical. Although approximately thirty manual pain instruments have been developed for neonatal pain diagnosis, most are complex, multifactorial, and geared toward research. The goals of this work are twofold: 1) to develop a new video dataset for automatic neonatal pain detection called iCOPEvid (infant Classification Of Pain Expressions videos), and 2) to present a classification system that sets a challenging comparison performance on this dataset. The iCOPEvid dataset contains 234 videos of 49 neonates experiencing a set of noxious stimuli, a period of rest, and an acute pain stimulus. From these videos 20 s segments are extracted and grouped into two classes: pain (49) and nopain (185), with the nopain video segments handpicked to produce a highly challenging dataset. An ensemble of twelve global and local descriptors with a Bag-of-Features approach is utilized to improve the performance of some new descriptors based on Gaussian of Local Descriptors (GOLD). The basic classifier used in the ensembles is the Support Vector Machine, and decisions are combined by sum rule. These results are compared with standard methods, some deep learning approaches, and 185 human assessments. Our best machine learning methods are shown to outperform the human judges.

The main goal of this chapter is to develop a system for automatic protein classification. Proteins are classified using CNNs trained on ImageNet, which are tuned using a set of multiview 2D images of 3D protein structures generated by Jmol, which is a 3D molecular graphics program. Jmol generates different types of protein visualizations that emphasize specific properties of a protein’s structure, such as a visualization that displays the backbone structure of the protein as a trace of the Cα atom. Different multiview protein visualizations are generated by uniformly rotating the protein structure around its central X, Y, and Z viewing axes to produce 125 images for each protein. This set of images is then used to fine-tune the pretrained CNNs. The proposed system is tested on two datasets with excellent results. The MATLAB code used in this chapter is available at https://github.com/LorisNanni.

Skin detectors play a crucial role in many applications: face localization, person tracking, objectionable content screening, etc. Skin detection is a complicated process that involves not only the development of apposite classifiers but also many ancillary methods, including techniques for data preprocessing and postprocessing. In this paper, a new postprocessing method is described that learns to select whether an image needs the application of various morphological sequences or a homogeneity function. The type of postprocessing method selected is learned based on categorizing the image into one of eleven predetermined classes. The novel postprocessing method presented here is evaluated on ten datasets recommended for fair comparisons that represent many skin detection applications. The results show that the new approach enhances the performance of the base classifiers and previous works based only on learning the most appropriate morphological sequences.

Bioimage classification is increasingly becoming more important in many biological studies including those that require accurate cell phenotype recognition, subcellular localization, and histopathological classification. In this paper, we present a new General Purpose (GenP) bioimage classification method that can be applied to a large range of classification problems. The GenP system we propose is an ensemble that combines multiple texture features (both handcrafted and learned descriptors) for superior and generalizable discriminative power. Our ensemble obtains a boosting of performance by combining local features, dense sampling features, and deep learning features. Each descriptor is used to train a different Support Vector Machine that is then combined by sum rule. We evaluate our method on a diverse set of bioimage classification tasks each represented by a benchmark database, including some of those available in the IICBU 2008 database. Each bioimage classification task represents a typical subcellular, cellular, and tissue level classification problem. Our evaluation on these datasets demonstrates that the proposed GenP bioimage ensemble obtains state-of-the-art performance without any ad-hoc dataset tuning of the parameters (thereby avoiding any risk of overfitting/overtraining). To reproduce the experiments reported in this paper, the MATLAB code of all the descriptors is available at https://github.com/LorisNanni and https://www.dropbox.com/s/bguw035yrqz0pwp/ElencoCode.docx?dl=0.

In this paper we present three different learning models for the automatic assessment of neonatal pain during clinical procedures. Given that few data are publicly available for the training and evaluation of these systems, we develop solutions that try to get the most out of the data that are available. To accomplish this, we choose a convolutional neural network (CNN) architecture as the discriminator with a reduced number of trainable parameters that are used efficiently. Furthermore, we develop two solutions based on the generative adversarial network (GAN) framework in order to improve discriminator power by transforming it into a multitask classifier and by training it with a combination of real and synthetic samples. Experimental results on the publicly available infant Classification Of Pain Expressions (iCOPE) database show superior results compared to previous works in the state of the art.

Pattern representation affects classification performance. Although discovering “universal” features that work for many classification problems is ideal, most representations are problem specific. In this paper, we improve the classification performance of a classifier system by transforming a one-dimensional input descriptor into a two-dimensional space so that effective texture extractors can be extracted to capture hidden data information. We develop two new methods for matrix representation where features are extracted that are more generalizable. The first method for generating a two-dimensional representation of patterns is based on trees, the objective being to reshape the feature vector into a matrix, and the second method performs mathematical operations to build a matrix representation. The proposed framework is then evaluated for its specific power on three medical problems. To evaluate generalizability, we compare the proposed approaches with several other baseline methods across some well-known benchmark datasets that reflect a diversity of classification problems. Not only does our approach show high robustness, but it also exhibits low sensitivity to parameters. When different approaches for transforming a vector into a matrix are combined with several texture descriptors, the resulting system often works well without requiring any ad-hoc optimization. The performance of the tested systems is compared by Wilcoxon signed rank test and Friedman's test.

A fundamental problem in computer vision is face detection. In this paper, an experimentally derived ensemble made by a set of six face detectors is presented that maximizes the number of true positives while simultaneously reducing the number of false positives produced by the ensemble. False positives are removed using different filtering steps based primarily on the characteristics of the depth map related to the subwindows of the whole image that contain candidate faces. A new filtering approach based on processing the image with different wavelets is also proposed here. The experimental results show that the applied filtering steps used in our best ensemble reduce the number of false positives without decreasing the detection rate. This finding is validated on a combined dataset composed of four others for a total of 549 images, including 614 upright frontal faces acquired in unconstrained environments. The dataset provides both 2D and depth data. For further validation, the proposed ensemble is tested on the well-known BioID benchmark dataset, where it obtains a 100% detection rate with an acceptable number of false positives.

In this work we propose a simple yet effective face detector that combines several face/eye detectors that possess different characteristics. Specifically, we report an extensive study for combining face/eye detectors that results in a final system we call FED that combines three face detectors that extract regions of candidate faces from an image with two approaches for eye detection: the enhanced Pictorial Structure (PS) model for coarse eye localization and a new approach proposed here (called PEC) that provides precise eye localization. PEC is an ensemble that utilizes three texture descriptors: multi-resolution local ternary patterns, local phase quantization descriptors, and patterns of oriented edge magnitudes. The extracted features are coupled with support vector machines trained on eye and non-eye samples to perform classification. The proposed framework for face detection could be considered an ad hoc integration of existing methods (the three face detectors and the PS coarse eye detector) that is combined with the proposed novel ensemble for precise eye localization (PEC). The aim of this approach is to maximize performance (not computation time). The quality of the proposed system is validated on three datasets (the well-known BioID and FERET datasets as well as a self-collected dataset). To the best of our knowledge, our system is one of the first fully automatic face detection approaches to obtain an accuracy of almost 100% on the BioID dataset (the most important benchmark dataset for frontal face detection) and 99.1% using the same dataset with only 12 false positives. A MATLAB version of our complete system for face detection can be downloaded from https://www.dei.unipd.it/node/2357.

Multiclassifier systems, the focus of this article, provide scientists and data professionals with powerful techniques for tackling complex datasets. The basic idea behind the multiclassifier approach is to average the decisions or hypotheses of a diverse group of classifiers in order to produce a better decision or hypothesis. As an introduction to our subject, we begin with a detailed examination of the canonic single-classifier system, as this provides the mathematical foundation needed in our presentation of multiclassifier systems. We then describe some important methods for constructing multiclassifier systems at all the levels mentioned above: the classifier level, the combination level, the data level, and the feature level. We thus end our overview of multiclassifier systems with a section that provides guidance for experimentally constructing general-purpose (GP) multiclassifier systems. Keywords: Machine learning; ensemble of classifiers

In this chapter, we present some variants of local phase quantization (LPQ), a novel texture descriptor that has been shown to perform well on a variety of classification tasks. After providing an extensive review of LPQ, we report experiments using several new LPQ derivatives obtained by varying LPQ parameters and by using a ternary rather than the binary encoding scheme. Multiple parameter sets are generated and each set is used to train a standard machine-learning classifier, a stand-alone support vector machine. The ensemble is then combined using the sum rule. Extensive experiments are conducted using six different datasets. Our method is compared along with the best state-of-the-art methods for solving each problem represented by the datasets. In each case, the best result is obtained using an ensemble with LPQ variants and ternary encoding. In this study, we also examine the distribution in the images of the most important bins of the LPQ histograms using Gabor filters. We find that incorporating this information into our best texture descriptor approach produces even better results.

This paper focuses on extensive testing of multimatcher methods for obtaining a system that is comparable with the state-of-the-art commercial matchers. Through extensive testing, we propose an ensemble method that uses minutiae, correlation-based and hybrid methods. To further improve our system we conduct experiments with matchers that are also based on different enhancement techniques, combining matcher scores by sum rule. Our results are validated on all four FVC2004 DBs and on the easier FVC2002 DB2. Finally, we study the fusion among the proposed methods with the competitor systems in FVC2004 competition. We find that our method improves the performance of the winner of FVC2004 competition. The MATLAB code for our experiments is freely available for downloading at bias.csr.unibo.it/nanni/Fconf.rar. Keywords

In this paper we make an extensive study of different methods for building ensembles of classifiers. We examine variants of ensemble methods that are based on perturbing features. We illustrate the power of using these variants by applying them to a number of different problems. We find that the best performing ensemble is obtained by combining an approach based on random subspace with a cluster based input decimated ensemble and the principal direction oracle. Compared with other state-of-the-art stand-alone classifiers and ensembles, this method consistently performed well across twelve diverse benchmark datasets. Another useful finding is that this approach does not need parameters to be carefully tuned for each dataset (in contrast to the fundamental importance of parameters tuning when using SVM and extreme learning machines), making our ensemble method well suited for practitioners since there is less risk of over-training. Another interesting finding is that random subspace can be coupled with several other ensemble methods to improve performance.

The most common method of handling automated cell phenotype image classification is to determine a common set of optimal features and then apply standard machine-learning algorithms to classify them. In this chapter, we use advanced methods for determining a set of optimized features for training an ensemble using random subspace with a set of Levenberg–Marquardt neural networks. The process requires that we first run several experiments to determine the individual features that offer the most information. The best performing features are then concatenated and used in the ensemble classification. Applying this approach, we have obtained an average accuracy of 97.4% using the three best benchmarks for this problem: the 2D HeLa dataset and both the endogenous and the transfected LOCATE mouse protein subcellular localization databases.