Francesco Camastra

A novel kernel method for clustering

Kernel methods are algorithms that, by replacing the inner product with an appropriate positive definite function, implicitly perform a nonlinear mapping of the input data into a high-dimensional feature space. In this paper, we present a kernel method for clustering inspired by the classical k-means algorithm in which each cluster is iteratively refined using a one-class support vector machine. Our method, which can be easily implemented, compares favorably with respect to popular clustering algorithms, like k-means, neural gas, and self-organizing maps, on a synthetic data set and three UCI real data benchmarks (IRIS data, Wisconsin breast cancer database, Spam database).

Data dimensionality estimation methods: a survey

In this paper, data dimensionality estimation methods are reviewed. The estimation of the dimensionality of a data set is a classical problem of pattern recognition. There are some good reviews (Algorithms for Clustering Data, Prentice-Hall, Englewood Cliffs, NJ, 1988) in literature but they do not include more recent developments based on fractal techniques and neural autoassociators. The aim of this paper is to provide an up-to-date survey of the dimensionality estimation methods of a data set, paying special attention to the fractal-based methods.

A SVM-based cursive character recognizer

This paper presents a cursive character recognizer, a crucial module in any cursive word recognition system based on a segmentation and recognition approach. The character classification is achieved by using support vector machines (SVMs) and a neural gas. The neural gas is used to verify whether lower and upper case version of a certain letter can be joined in a single class or not. Once this is done for every letter, the character recognition is performed by SVMs. A database of 57293 characters was used to train and test the cursive character recognizer. SVMs compare notably better, in terms of recognition rates, with popular neural classifiers, such as learning vector quantization and multi-layer-perceptron. SVM recognition rate is among the highest presented in the literature for cursive character recognition.

Cursive character recognition by learning vector quantization

This paper presents a cursive character recognizer embedded in an off-line cursive script recognition system. The recognizer is composed of two modules: the first one is a feature extractor, the second one a learning vector quantizer. The selected feature set was compared to Zernike polynomials using the same classifier. Experiments are reported on a database of about 49,000 isolated characters.

A fuzzy decision system for genetically modified plant environmental risk assessment using Mamdani inference

A fuzzy system for environmental risk assessment of genetically modified plants is described.The fuzzy system is based on Mamdani inference.The Fuzzy System Risk Assessments have been validated on real world trial case studies.The system decisions have been considered coherent and consistent by human experts. Environmental risk assessment (ERA) of the deliberate release of genetically modified plants (GMPs) is currently performed by human experts on the basis of own personal experience and knowledge. In this paper we describe a fuzzy decision system (FDS) for the ERA of GMPs, based on Mamdani fuzzy inference. The risk assessment in the FDS is obtained by using a fuzzy inference system (FIS), performed using jFuzzyLogic library. The FDS permits obtaining an evaluation process for the identification of potential impacts that can achieve one or more receptors through a set of migration paths. The decisions derived by FDS have been validated on real world cases by the human experts that are in charge of ERA. They have confirmed the reliability and correctness of the fuzzy system decisions.

Real-time hand gesture recognition using a color glove

This paper presents a real-time hand gesture recognizer based on a color glove. The recognizer is formed by three modules. The first module, fed by the frame acquired by a webcam, identifies the hand image in the scene. The second module, a feature extractor, represents the image by a nine-dimensional feature vector. The third module, the classifier, is performed by means of Learning Vector Quantization. The recognizer, tested on a dataset of 907 hand gestures, has shown very high recognition rate.

Combining Neural Gas and Learning Vector Quantization for Cursive Character Recognition

This paper presents a cursive character recognizer, a crucial module in any Cursive Script Recognition system based on a segmentation and recognition approach. n nThe character classification is achieved by combining the use of neural gas (NG) and learning vector quantization (LVQ). NG is used to verify whether lower and upper case version of a certain letter can be joined in a single class or not. Once this is done for every letter, it is possible to find an optimal number of classes maximizing the accuracy of the LVQ classifier. n nA database of 58000 characters was used to train and test the models. The performance obtained is among the highest presented in the literature for the recognition of cursive characters.

Intrinsic Dimension Estimation of Data: An Approach Based on Grassberger–Procaccia's Algorithm

In this paper the problem of estimating the intrinsic dimension of a data set is investigated. An approach based on the Grassberger–Procaccias algorithm has been studied. Since this algorithm does not yield accurate measures in high-dimensional data sets, an empirical procedure has been developed. Grassberger–Procaccias algorithm was tested on two different benchmarks and was compared to a TRN-based method.

Offline Cursive Character Challenge: a New Benchmark for Machine Learning and Pattern Recognition Algorithms.

Cursive character recognition is a challenging task due to high variability and intrinsic ambiguity of cursive letters. This paper presents C-Cube (Cursive Character Challenge), a new public-domain cursive character database. C-Cube contains 57293 cursive characters manually extracted from cursive handwritten words, including both upper and lower case versions of each letter. The database can be downoloaded from the Web and it provides predefined experimental protocols in order to compare rigorously the results obtained by different researchers

Machine Learning for Audio, Image and Video Analysis

The past decade has witnessed an explosion of digital information and a phenomenal growth in the popularity of audio, image, and video multimedia. Due to the advancements in processing power and the proliferation of the Internet, people can easily capture, store, transmit, and share audio, image, and video content. However, efficient and effective indexing and retrieval of such data accumulated over time is still a formidable challenge. Researchers in industry and academia have spent a tremendous effort on developing sophisticated systems for processing and understanding this new information, at the core of which always lie machine-learning techniques. As a very broad research field, machine learning covers a broad set of areas ranging from uncertainty analysis to kernel methods. Although it closely interacts with other fields such as statistics, signal processing, and pattern recognition, it is almost impossible for the researchers in those areas to understand every detail of the state-of-the-art machine-learning techniques. While most of the books on machine learning cover classic techniques such as neural networks or support vector machines, few of them emphasize the recent advancements and their applications to audio, image, and video analysis. The book Machine Learning for Audio, Image and Video Analysis intends to fill this gap by bringing to its readers the latest developments in this fast-growing field. The book consists of an introduction, three main parts total of 13 chapters , and four appendices. The introduction explains how readers with various backgrounds can benefit from studying this book and provides the ABCs about acquisition and processing of audio and visual information for beginners. The appendices also prepare supplemental material for readers who lack the related statistical and signal-processing background. For experienced researchers, the recent advancements can be found in Part II, where classic machine-learning techniques are discussed along with their state-of-the-art developments. For practitioners, the authors analyze three typical applications to provide a sense of how machine-learning techniques can be applied to understand audio/video data. Part I contains a brief description of how the human biological system perceives audio and video signals and how these signals are captured and digitized in a format that is amenable to computer processing. Chapters 2 and 3 further introduce the audio/video representation and coding standards without getting into too much detail. Besides color, texture, and shape, it is my opinion that local descriptors such as scale-invariant feature transform SIFT should be introduced at this point because they have demonstrated the strength of these “bag of words” representations in many applications. Widely used machine-learning techniques are the focus of Chapters 4 to 11 in Part II. Chapters 4, 5, and 7 introduce the general objectives and approaches of machine learning and the means to evaluate its performance from a statistical point of view. Chapter 5 introduces the Bayesian decision theory, which leads to further investigation of Markovian models in Chapter 10. Kernel methods are discussed in detail in Chapter 9. It is worth mentioning that this book, unlike most other books in this field, not only introduces a few widely used techniques in audio and image analysis, but also discusses the latest advancements in the field. For example, most books would touch the surface of support vector machines SVM by introducing the original two-class SVM, yet Chapter 9 of this book goes one step further to discuss sequential minimal optimization, a powerful multiclass extension of SVM, which is more appealing in practice. Chapters 6 and 11 are concerned with clustering and dimension reduction via unsupervised learning. Specifically, Chapter 11 introduces several manifold learning techniques, such as locally linear embedding LLE and ISOMAP, which are particularly useful in handling nonlinear data in audio and video processing. Chapter 8 combines the discussion of classic neural networks and ensemble methods. My personal view is that ensemble methods, such as AdaBoost and random forest, deserve an independent chapter because of their outstanding reported performance on many applications and benchmark data sets. Similarly, the “topic model” methods, such as probabilistic latent semantic indexing pLSI and latent dirichlet analysis LDA , should be added to this book to reflect the current research trend in analyzing text and visual data. Part III showcases three applications of machine-learning techniques, namely speech and handwriting recognition, automatic face recognition, and video segmentation and keyframe extraction. In Chapter 13, the authors discuss the automatic facerecognition system. However, face image localization—one of the core problems in this application—doesn’t seem to get enough attention. It would be desirable to add a section about the boosted cascade method proposed by Viola and Jones, which is one of the most successful machine-learning applications in image analysis. In addition, it is better to point out that the eigenface and its variants can suppress a lot of the luminance variation of the face images by removing the eigenvectors corresponding to the three largest eigenvalues. There are several things that are unique in this book. In some chapters, the problem sections are included to challenge the readers to understand the discussed methods or apply them to solve some sample problems. Distinct from other books, it also points out several public software packages and benchmark data sets that encourage the reader to have a hands-on experience on how machine-learning techniques work to analyze audio and visual content. Its comprehensive coverage on recent development in this research area makes it easy for experienced researchers to further explore the latest techniques.