Search

We consider the problem of recognizing face or object when only single training image per class is available, which is typically encountered in law enforcement, passport or identification card verification, etc. In such cases, many discriminant subspace methods such as Linear Discriminant Analysis (LDA) fail because of the non-existence of intra-class variation. In this paper, we propose a novel framework called 2-Dimensional Kernel PCA (2D-KPCA) for face or object recognition from a single image. In contrast to conventional KPCA, 2D-KPCA is based on 2D image matrices and hence can effectively utilize the intrinsic spatial structure information of the images. On the other hand, in contrast to 2D-PCA, 2D-KPCA is capable of capturing part of the higher-order statistics information. Moreover, this paper reveals that the current 2D-PCA algorithm and its many variants consider only the row information or column information, which has not fully exploited the information contained in the image matrices. So, besides proposing the unilateral 2D-KPCA, this paper also proposes the bilateral 2D-KPCA which could exploit more information concealed in the image matrices Furthermore, some approximation techniques are developed for improving the computational efficiency. Experimental results on the FERET face database and the COIL-20 object database show that: 1) the performance of KPCA is not necessarily better than that of PCA; 2) 2D-KPCA almost always outperforms 2D-PCA significantly; 3) the kernel methods are more appropriate on 2D pattern than on 1D patterns. [ABSTRACT FROM AUTHOR]

In this paper we propose a method based on a graph-theoretical cluster analysis for automatically finding and classifying clusters of microcalcifications in mammographic images, starting from the output of a microcalcification detection phase. This method does not require the user to provide either the expected number of clusters or any threshold values, often with no clear physical meaning, as other algorithms do. The proposed approach has been tested on a standard database of 40 mammographic images and has demonstrated to be very effective, even when the detection phase gives rise to several false positives. [ABSTRACT FROM AUTHOR]

The aim of this paper is to present an efficient distance between n-dimensional histograms. Some image classification or image retrieval techniques use the distance between histograms as a first step of the classification process. For this reason, some algorithms that find the distance between histograms have been proposed in the literature. Nevertheless, most of this research has been applied on one-dimensional histograms due to the computation of a distance between multi-dimensional histograms is very expensive. In this paper, we present an efficient method to compare multi-dimensional histograms in O(2z), where z represents the number of bins. Results show a huge reduction of the time consuming with no recognition-ratio reduction. [ABSTRACT FROM AUTHOR]

In this paper, we propose a block-based real-time people counting system that can be used in various environments including shopping mall entrances, elevators and escalators. The main contributions of this paper are robust background subtraction, the block-based decision method and real-time processing. For robust background subtraction obtained from a number of image sequences, we used a mixture of K Gaussian. The block-based decision method was used to determine the size of the given objects (moving people) in each block. We divided the images into 72 blocks and trained the mean and variance values of the specific objects in each block. This was done in order to provide real-time processing for up to 4 channels. Finally, we analyzed various actions that can occur with moving people in real world environments. [ABSTRACT FROM AUTHOR]

We present the ubiquitous intelligent sensing system for a smart home in this paper. A smart home is intelligent space that studies patterns of home contexts that is acquired in a home, and provides automatic home services for the human. The ubiquitous intelligent sensing system acquires seven sensing contexts from four sensor devices. We utilize association rules of data mining and linear support machine to analyze context patterns of seven contexts. Also, we analyze stress rates of the human through the HRV pattern of the ECG. If the human is suffering from stress, the ubiquitous intelligent sensing system provides home service to reduce one's stress. In this paper, we present the architecture and algorithms of the ubiquitous intelligent sensing system. We present the management toolkit to control the ubiquitous intelligent sensing system, and show implementation results of the smart home using the ubiquitous intelligent sensing system. [ABSTRACT FROM AUTHOR]

This paper describes a feature extraction method for realtime surveillance. Eigenspace models are a convenient way to represent set of images with widespread applications. In the traditional approach to calculate these eigenspace models, known as batch PCA method, model must capture all the images needed to build the internal representation. This approach has some drawbacks. Since the entire set of images is necessary, it is impossible to make the model build an internal representation while exploring a new person. Updating of the existing eigenspace is only possible when all the images must be kept in order to update the eigenspace, requiring a lot of storage capability. In this paper we propose a method that allows for incremental eigenspace update method by incremental kernel PCA for realtime surveillance. Experimental results indicate that accuracy of proposed method is comparable to batch KPCA and outperforms than APEX. Furthermore proposed method has efficiency in memory requirement compared to KPCA. [ABSTRACT FROM AUTHOR]

We re-visit the age-old problem of estimating the parameters of a distribution from its observations. Traditionally, scientists and statisticians have attempted to obtain strong estimates by "extracting" the information contained in the observations taken as a set. However, generally speaking, the information contained in the sequence in which the observations have appeared, has been ignored - i.e., except to consider dependence information as in the case of Markov models and n-gram statistics. In this paper, we present results which, to the best of our knowledge, are the first reported results, which consider how estimation can be enhanced by utilizing both the information in the observations and in their sequence of appearance. The strategy, known as Sequence Based Estimation (SBE) works as follows. We first quickly allude to the results pertaining to computing the Maximum Likelihood Estimates (MLE) of the data when the samples are taken individually. We then derive the corresponding MLE results when the samples are taken two-at-a-time, and then extend these for the cases when they are processed three-at-a-time, four-at-a-time etc. In each case, we also experimentally demonstrate the convergence of the corresponding estimates. We then suggest various avenues for future research, including those by which these estimates can be fused to yield a superior overall cumulative estimate of the parameter of the distribution. We believe that our new estimates have great potential for practitioners, especially when the cardinality of the observation set is small. [ABSTRACT FROM AUTHOR]

Fisher's Linear Discriminant Analysis (LDA) is a traditional dimensionality reduction method that has been proven to be successful for decades. Numerous variants, such as the Kernel-based Fisher Discriminant Analysis (KFDA) have been proposed to enhance the LDA's power for nonlinear discriminants. Though effective, the KFDA is computationally expensive, since the complexity increases with the size of the data set. In this paper, we suggest a novel strategy to enhance the computation for an entire family of KFDA's. Rather than invoke the KFDA for the entire data set, we advocate that the data be first reduced into a smaller representative subset using a Prototype Reduction Scheme (PRS), and that dimensionality reduction be achieved by invoking a KFDA on this reduced data set. In this way data points which are ineffective in the dimension reduction and classification can be eliminated to obtain a significantly reduced kernel matrix, K, without degrading the performance. Our experimental results demonstrate that the proposed mechanism dramatically reduces the computation time without sacrificing the classification accuracy for artificial and real-life data sets. [ABSTRACT FROM AUTHOR]

Sometimes novel or outlier data has to be detected. The outliers may indicate some interesting rare event, or they should be disregarded because they cannot be reliably processed further. In the ideal case that the objects are represented by very good features, the genuine data forms a compact cluster and a good outlier measure is the distance to the cluster center. This paper proposes three new formulations to find a good cluster center together with an optimized ℓp-distance measure. Experiments show that for some real world datasets very good classification results are obtained and that, more specifically, the ℓ1-distance is particularly suited for datasets containing discrete feature values. Keywords: one-class classification, outlier detection, robustness, ℓp-ball. [ABSTRACT FROM AUTHOR]

As radar backscatter values for oil slicks are very similar to backscatter values for very calm sea areas and other ocean phenomena, dark areas in Synthetic Aperture Radar (SAR) imagery tend to be misinterpreted. In this paper three feature sets are used to identify the oil slicks in SAR images. These images are submitted to different MLP architectures to verify the separability performance over each feature set. This analysis is very suitable for remote sensing of environment applications concerning marine oil pollution. The estimated resulting performance points out which feature set is the best suitable for the suggested application. [ABSTRACT FROM AUTHOR]

In this paper, we present an audio scheme protective of copyright protection using information hiding. We propose visually recognizable binary image and text information as watermark (copyright) information embedded in audio signal. Cepstrum representation of audio can be shown to be very robust to a wide range of attacks. We apply SMM(statistical mean manipulation) theory in embedding image watermarking, and address attacks against lossy audio compression like MP3, white Gaussian noise and so on. A blind detection watermarking can be realized with the proposed scheme. Keywords: copyright protection, information hiding, watermark, statistical mean manipulation, cepstrum domain. [ABSTRACT FROM AUTHOR]

Based on template facial features and image segmentation, this paper demonstrates a novel method for automatic detection of eyes in grayscale still images. A decision model of eye location is instituted by the priori knowledge of template facial features. After roughly detection of face, we apply three steps for system to locate eyes. Firstly, the Bayesian eye detector is used to find eye patterns in the upper region of the face image. This vector based Bayesian classifier adopts Haar transform as vectorize because we know that is robust at illumination variation. Secondly, merging and arbitration strategy are applied. It can manage variations of around eye regions due to spectacle rims or eye brows. Finally, Gaussian-projection function can locate robust precision eye position. The experimental results show that the proposed method can achieve higher performance at any test data. [ABSTRACT FROM AUTHOR]

For an application problem, there may be multiple databases, and each database may not contain complete variables or attributes, that is, some variables are observed but some others are missing. Further, data of a database may be collected conditionally on some designed variables. In this paper, we discuss problems related to data mining from such multiple databases. We propose an approach for detecting identifiability of a joint distribution from multiple databases. For an identifiable joint distribution, we further present the expectation-maximization (EM) algorithm for calculating the maximum likelihood estimates (MLEs) of the joint distribution. [ABSTRACT FROM AUTHOR]

In this paper we present a novel algorithm for anchor shot detection (ASD). ASD is a fundamental step for segmenting news video into stories that is among key issues for achieving efficient treatment of news-based digital libraries. The proposed algorithm creates a set of audio/video templates of anchorperson shots in an unsupervised way, then classifies shots by comparing them to the templates. Audio similarity is evaluated by means of a new index and helps to achieve better performance than a pure video approach. The method has been tested on a wide database and compared with other state-of-the-art algorithms, demonstrating its effectiveness with respect to them. [ABSTRACT FROM AUTHOR]

Editing allows the selection of a representative subset of prototypes among the training sample to improve the performance of a classification task. The Wilson's editing algorithm was the first proposal and then a great variety of new editing techniques have been proposed based on it. This algorithm consists on the elimination of prototypes in the training set that are misclassified using the k-NN rule. From such editing scheme, a general editing procedure can be straightforward derived, where any classifier beyond k-NN can be used. In this paper, we analyze the behavior of this general editing procedure combined with 3 different neighborhood-based classification rules, including k-NN. The results reveal better performances of the 2 other techniques with respect to k-NN in most of cases. Keywords: Pattern recognition, classification, nearest neighbor, prototype selection, editing. [ABSTRACT FROM AUTHOR]

In this paper, we propose an approach to image content recognition that exploits the benefits of different image representations to associate meaning with images. We choose classifiers based on global appearance, scene structure and region type occurrence, and define confidence measures on their output. The resulting posterior probabilities of the classifiers are combined in a Bayesian framework. We show that this method leads to a robust and efficient system that contributes to reducing the semantic gap between low level image features and higher level image descriptions. [ABSTRACT FROM AUTHOR]

In this paper we address the problem of estimating the parameters of a Gaussian mixture model. Although the EM (Expectation-Maximization) algorithm yields the maximum-likelihood solution it requires a careful initialization of the parameters and the optimal number of kernels in the mixture may be unknown beforehand. We propose a criterion based on the entropy of the pdf (probability density function) associated to each kernel to measure the quality of a given mixture model. Two different methods for estimating Shannon entropy are proposed and a modification of the classical EM algorithm to find the optimal number of kernels in the mixture is presented. We test our algorithm in probability density estimation, pattern recognition and color image segmentation. [ABSTRACT FROM AUTHOR]

In this paper we propose a general framework for analysing the diversity of ensembles of word sequence recognition systems. The goal of the framework is to enable the application of any diversity measure developed for standard multi-class classification problems to ensembles of word sequence recognisers. Experiments with several diversity measures are conducted on artificial as well as on real world data and show the effectiveness of the proposed approach. [ABSTRACT FROM AUTHOR]

We propose probabilistic framework for analysis of inaccuracies due to feature selection (FS) when flawed estimates of performance of feature subsets are utilized. The approach is based on analysis of random search FS procedure and postulation that joint distribution of true and estimated classification errors is known a priori. We derive expected values for the FS bias, a difference between actual classification error after FS and classification error if ideal FS is performed according to exact estimates. The increase in true classification error due to inaccurate FS is comparable or even exceeds a training bias, a difference between generalization and Bayes errors. We have shown that there exists overfitting phenomenon in feature selection, entitled in this paper as feature over-selection. The effects of feature over-selection could be reduced if FS would be performed on basis of positional statistics. Theoretical results are supported by experiments carried out on simulated Gaussian data, as well as on high dimensional microarray gene expression data. [ABSTRACT FROM AUTHOR]

This paper proposes a Sequential Monte Carlo (SMC) learning algorithm for Bayesian probability distributions that describe model parameters in a video face recognition system based on deformable template matching. The new algorithm achieves significantly improved robustness of recognition against facial expressions and speech movements by comparison with a baseline batch MCMC (Markov Chain Monte Carlo) algorithm, at no additional computational cost. Experimental results demonstrate the effectiveness and computational efficiency of the new algorithm. [ABSTRACT FROM AUTHOR]

This paper presents an object recognition method based on recursive neural networks (RNNs) and multiresolution trees (MRTs). MRTs are a novel hierarchical structure proposed to represent both the set of homogeneous regions in which images can be divided and the evolution of the segmentation process performed to determine such regions. Moreover, knowing the optimal number of regions that should be extracted from the images is not critical for the construction of MRTs, that are also invariant w.r.t. rotations and translations. A set of experiments was performed on a subset of the Caltech benchmark database, comparing the performances of the MRT and directed acyclic graph (DAG) representations. The results obtained by the proposed object detection technique are also very promising in comparison with other state-of-the-art approaches available in the literature. [ABSTRACT FROM AUTHOR]

Considering the classification problem in which class priors or misallocation costs are not known precisely, receiver operator characteristic (ROC) analysis has become a standard tool in pattern recognition for obtaining integrated performance measures to cope with the uncertainty. Similarly, in situations in which priors may vary in application, the ROC can be used to inspect performance over the expected range of variation. In this paper we argue that even though measures such as the area under the ROC (AUC) are useful in obtaining an integrated performance measure independent of the priors, it may also be important to incorporate the sensitivity across the expected prior-range. We show that a classifier may result in a good AUC score, but a poor (large) prior sensitivity, which may be undesirable. A methodology is proposed that combines both accuracy and sensitivity, providing a new model selection criterion that is relevant to certain problems. Experiments show that incorporating sensitivity is very important in some realistic scenarios, leading to better model selection in some cases. [ABSTRACT FROM AUTHOR]

In this paper, we proposed a method to speed up the test phase of SVM based on Feature Vector Selection method (FVS). In the method, the support vectors (SVs) appeared in the decision function of SVM are replaced with some feature vectors (FVs) which are selected from support vectors by FVS method. Since it is a subset of SVs set, the size of FVs set is normally smaller than that of the SVs set, therefore the decision process of SVM is speeded up. Experiments on 12 datasets of IDA show that the number of SVs can be reduced from 20% to 99% with only a slight increase on the error rate of SVM by the proposed method. The trade-off between the generalization ability of obtained SVM and the speedup ability of the proposed method can be easily controlled by one parameter. [ABSTRACT FROM AUTHOR]

This paper presents a context driven segmentation and recognition method for handwritten Chinese characters. We follow a split-merge technique in character segmentation. In this process, a Chinese text line is first pre-segmented into a sequence of radicals, which are then merged according to a cost function combining both recognition confidence and contextual cost. Two strategies are also proposed for implementation: bi-gram based merging and lexicon driven merging. In the former one, we generate a set of merging paths which are then evaluated by Viterbi algorithm. The radicals' best merging method is given by the path with the highest score. In the latter strategy, a lexicon is preset and compared with the radicals to determine both radicals' merging and candidate character selection. Experiments show that contextual information plays a crucial role in Chinese character segmentation and could obviously improve the segmentation and recognition results. [ABSTRACT FROM AUTHOR]

This paper explores the use of commute-time preserving embedding as means of data-clustering. Commute time is a measure of the time taken for a random walk to set-out and return between a pair of nodes on a graph. It may be computed from the spectrum of the Laplacian matrix. Since the commute time is averaged over all potential paths between a pair of nodes, it is potentially robust to variations in graph structure due to edge insertions or deletions. Here we demonstrate how nodes of a graph can be embedded in a vector space in a manner that preserves commute time. We present a number of important properties of the embedding. We experiment with the method for separating object motions in image sequences. [ABSTRACT FROM AUTHOR]

The aim of this paper is to present a new method to compare modulo histograms. In these histograms, the type of elements are cyclic, for instance, the hue in colour images. The main advantage is that there is an important time-complexity reduction respect the methods presented before. The distance between histograms that we present is defined on a structure called signature, which is a lossless representation of histograms. We show that the computational cost of our distance is O($\mathcal{}z'$2), being $\mathcal{}z'$ the number of non-empty bins of the histograms. The computational cost of the algorithms presented in the literature depends on the number of bins of the histograms. In most of the applications, the obtained histograms are sparse, then considering only the non-empty bins makes the time consuming of the comparison drastically decrease. The distance and algorithms presented in this paper are experimentally validated on the comparison of images obtained from public databases. [ABSTRACT FROM AUTHOR]

Semi-automatic parking system is a driver convenience system automating steering control required during parking operation. This paper proposes novel monocular-vision based target parking-slot recognition by recognizing parking-slot markings when driver designates a seed-point inside the target parking-slot with touch screen. Proposed method compensates the distortion of fisheye lens and constructs a bird's eye view image using homography. Because adjacent vehicles are projected along the outward direction from camera in the bird's eye view image, if marking line-segment distinguishing parking-slots from roadway and front-ends of marking line-segments dividing parking-slots are observed, proposed method successfully recognizes the target parking-slot marking. Directional intensity gradient, utilizing the width of marking line-segment and the direction of seed-point with respect to camera position as a prior knowledge, can detect marking line-segments irrespective of noise and illumination variation. Making efficient use of the structure of parking-slot markings in the bird's eye view image, proposed method simply recognizes the target parking-slot marking. It is validated by experiments that proposed method can successfully recognize target parking-slot under various situations and illumination conditions. [ABSTRACT FROM AUTHOR]

Fingerprint recognition systems are widely used in the field of biometrics. Many existing fingerprint sensors acquire fingerprint images as the user's fingerprint is contacted on a solid flat sensor. Because of this contact, input images from the same finger can be quite different and there are latent fingerprint issues that can lead to forgery and hygienic problems. For these reasons, a touchless fingerprint recognition system has been investigated, in which a fingerprint image can be captured without contact. While this system can solve the problems which arise through contact of the user's finger, other challenges emerge, for example, low ridge-valley contrast, and 3D to 2D image mapping. In this paper we discuss both the disadvantages and the advantages of touchless fingerprint systems and introduce the hardware and algorithm approach to solve the problems. We describe the structure of illuminator and the wavelength of light to acquire a high contrast fingerprint images. To solve the problem of 3D to 2D image mapping, we describe the method to remove the strong view difference fingerprint images. Experiments show that the touchless fingerprint system has better performance than the conventional touch based system. [ABSTRACT FROM AUTHOR]

Camera motion classification is an important issue in content-based video retrieval. In this paper, a robust and hierarchical camera motion classification approach is proposed. As the Support Vector Machine (SVM) has a very good learning capacity with limited sample set and does not require any heuristic parameter, the SVM is first employed to classify camera motions into translation and non-translation motions in preliminary classification. In this step, four features are extracted as input of the SVM. Then, zoom and rotation motions are further classified by analyzing the motion vectors' distribution. And the directions of translation motions are also identified. The experimental results show that the proposed approach achieves a good performance. [ABSTRACT FROM AUTHOR]

In this paper, we explore how the trace of the heat kernel can be used to characterise graphs for the purposes of measuring similarity and clustering. The heat-kernel is the solution of the heat-equation and may be computed by exponentiating the Laplacian eigensystem with time. To characterise the shape of the heat-kernel trace we use the zeta-function, which is found by exponentiating and summing the reciprocals of the Laplacian eigenvalues. From the Mellin transform, it follows that the zeta-function is the moment generating function of the heat-kernel trace. We explore the use of the heat-kernel moments as a means of characterising graph structure for the purposes of clustering. Experiments with the COIL and Oxford-Caltech databases reveal the effectiveness of the representation. [ABSTRACT FROM AUTHOR]

A common way of expressing string similarity in structural pattern recognition is the edit distance. It allows one to apply the kNN rule in order to classify a set of strings. However, compared to the wide range of elaborated classifiers known from statistical pattern recognition, this is only a very basic method. In the present paper we propose a method for transforming strings into n-dimensional real vector spaces based on prototype selection. This allows us to subsequently classify the transformed strings with more sophisticated classifiers, such as support vector machine and other kernel based methods. In a number of experiments, we show that the recognition rate can be significantly improved by means of this procedure. [ABSTRACT FROM AUTHOR]

Currently there has been much interested in developing the watermarking for 3D graphic data of mesh model or NURBS. However, the watermarking technique based on 3D CAD drawing leaves something to be desired. This paper proposes a watermarking technique for 3D design drawing using the components of Line, 3DFACE and ARC based on vertex that prevent the infringement of copyright from unlawfulness reproductions and distribution. By experimental result, we confirmed the invisibility of embedded watermarks as well as the robustness in geometrical attacks and file conversion to DWG, DXF, DWT and DWS. Keywords: 3D CAD Drawing, Watermarking. [ABSTRACT FROM AUTHOR]

Random walk kernels in conjunction with Support Vector Machines are powerful methods for error-tolerant graph matching. Because of their local definition, however, the applicability of random walk kernels strongly depends on the characteristics of the underlying graph representation. In this paper, we describe a simple extension to the standard random walk kernel based on graph edit distance. The idea is to include global matching information in the local similarity evaluation of random walks in graphs. The proposed extension allows us to improve the performance of the random walk kernel significantly. We present an experimental evaluation of our method on three difficult graph datasets. [ABSTRACT FROM AUTHOR]

Graph edit distance is one of the most flexible mechanisms for error-tolerant graph matching. Its key advantage is that edit distance is applicable to unconstrained attributed graphs and can be tailored to a wide variety of applications by means of specific edit cost functions. Its computational complexity, however, is exponential in the number of vertices, which means that edit distance is feasible for small graphs only. In this paper, we propose two simple, but effective modifications of a standard edit distance algorithm that allow us to suboptimally compute edit distance in a faster way. In experiments on real data, we demonstrate the resulting speedup and show that classification accuracy is mostly not affected. The suboptimality of our methods mainly results in larger inter-class distances, while intra-class distances remain low, which makes the proposed methods very well applicable to distance-based graph classification. [ABSTRACT FROM AUTHOR]

This paper presents a novel approach of distinguishing in-vocabulary (IV) words and out-of-vocabulary (OOV) words by using confidence score-based unsupervised incremental adaptation. The unsupervised adaptation uses Viterbi decode results which have high confidence scores to adjust new acoustic models. The adjusted acoustic models can award IV words and punish OOV words in confidence score, thus obtain the goal of separating IV and OOV words. Our Automatic Speech Recognition Laboratory has developed a Speech Recognition Developer Kit (SRDK) which serves as a baseline system for different speech recognition tasks. Experiments conducted on the SRDK system have proved that this method can achieve a rise over 41% in OOV words detection rate (from 68% to 96%) at the same cost of a false alarm (taken IV words as OOV words) rate of 10%. This method also obtains a rise over 11% in correct acceptance rate (from 88% to 98%) at the same cost of a false acceptance rate of 20%. [ABSTRACT FROM AUTHOR]

The combination of classifiers is an established technique to improve the classification performance. The combination rules proposed up to now generally try to decrease the classification error rate, which is a performance measure not suitable in many real situations and particularly when dealing with two class problems. In this case, a good alternative is given by the Area under the Receiver Operating Characteristic curve (AUC). This paper presents a method for the linear combination of two-class classifiers aimed at maximizing the AUC. The effectiveness of the approach has been confirmed by the tests performed on standard datasets. [ABSTRACT FROM AUTHOR]

This paper addresses the problem of facial expressions recognition using principal component analysis and independent component analysis onto dimension of the emotion. To reflect well the changes in facial expressions, a representation based on principal component analysis (PCA) excluded the first 2 principal components is presented, ICA representation from this PCA representation is developed. Facial expression performance in two dimensional structure was significant 90.9% in pleasure/displeasure dimension and 66.6% in the arousal/sleep dimension. The findings indicate that the two dimensional structure of emotion may reflect various emotion states as a stabled structure for the facial expression recognition. [ABSTRACT FROM AUTHOR]

Although fingerprint verification systems have attained a good performance, researchers recently pointed out their weakness under fraudulent attacks by fake fingers. In fact, the acquisition sensor can be deceived by fake fingerprints created with liquid silicon rubber. Among the solutions to this problem, the software-based ones are the cheapest and less intrusive. They use feature vectors made up of measures extracted from one or multiple impressions (static measures) or multiple frames (dynamic measures) of the same finger in order to distinguish live and fake fingers. In this paper, we jointly use both static and dynamic features and report an experimental investigation aimed to compare them and select the most effective ones. [ABSTRACT FROM AUTHOR]

Statistical learning algorithms often rely on the Euclidean distance. In practice, non-Euclidean or non-metric dissimilarity measures may arise when contours, spectra or shapes are compared by edit distances or as a consequence of robust object matching [1,2]. It is an open issue whether such measures are advantageous for statistical learning or whether they should be constrained to obey the metric axioms. The k-nearest neighbor (NN) rule is widely applied to general dissimilarity data as the most natural approach. Alternative methods exist that embed such data into suitable representation spaces in which statistical classifiers are constructed [3]. In this paper, we investigate the relation between non-Euclidean aspects of dissimilarity data and the classification performance of the direct NN rule and some classifiers trained in representation spaces. This is evaluated on a parameterized family of edit distances, in which parameter values control the strength of non-Euclidean behavior. Our finding is that the discriminative power of this measure increases with increasing non-Euclidean and non-metric aspects until a certain optimum is reached. The conclusion is that statistical classifiers perform well and the optimal values of the parameters characterize a non-Euclidean and somewhat non-metric measure. [ABSTRACT FROM AUTHOR]

Learning based on kernel machines is widely known as a powerful tool for various fields of information science such as pattern recognition and regression estimation. The efficacy of the model in kernel machines depends on the distance between the unknown true function and the linear subspace, specified by the training data set, of the reproducing kernel Hilbert space corresponding to an adopted kernel. In this paper, we propose a framework for the model selection of kernel-based learning machines, incorporating a class of kernels with an invariant metric. [ABSTRACT FROM AUTHOR]

Clustering in data mining is a discovery process that groups a set of data so as to maximize the intra-cluster similarity and to minimize the inter-cluster similarity. Clustering becomes more challenging when data are categorical and the amount of available memory is less than the size of the data set. In this paper, we introduce CBC (Clustering Based on Compressed Data), an extension of the Birch algorithm whose main characteristics refer to the fact that it can be especially suitable for very large databases and it can work both with categorical attributes and mixed features. Effectiveness and performance of the CBC procedure were compared with those of the well-known K-modes clustering algorithm, demonstrating that the CBC summary process does not affect the final clustering, while execution times can be drastically lessened. [ABSTRACT FROM AUTHOR]

Recently, the importance of incremental learning in changing environments has been acknowledged. This paper proposes a new ensemble learning method based on two level hypothesis tests for incremental learning in concept changing environments. We analyze the classification error as a stochastic variable, and introduce hypothesis test as mechanism for adaptively selecting classifiers. Hypothesis tests are used to distinguish between useful and useless individual classifiers and to identify classifier to be updated. Classifiers deemed as useful by the hypothesis test are integrated to form the final prediction. Experiments with simulated concept changing scenarios show that the proposed method could adaptively choose proper classifiers and adapt quickly to different concept changes to maintain its performance level. [ABSTRACT FROM AUTHOR]

In this paper we present an unsupervised algorithm to select the most adequate grouping of regions in an image using a hierarchical clustering scheme. Then, we introduce an optimisation approach for the whole process. The grouping method presented is based on the maximisation of a measure that represents the perceptual decision. The whole strategy takes profit from a hierarchical clustering to find a maximum of the proposed criterion. The algorithm has been used to segment real images as well as multispectral images achieving very accurate results on this task. [ABSTRACT FROM AUTHOR]

Similarity based classification methods use positive semi-definite (PSD) similarity matrices. When several data representations (or metrics) are available, they should be combined to build a single similarity matrix. Often the resulting combination is an indefinite matrix and can not be used to train the classifier. In this paper we introduce new methods to build a PSD matrix from an indefinite matrix. The obtained matrices are used as input kernels to train Support Vector Machines (SVMs) for classification tasks. Experimental results on artificial and real data sets are reported. [ABSTRACT FROM AUTHOR]

This paper presents a system to recognise cursive Arabic typewritten text. The system is built using the Hidden Markov Model Toolkit (HTK) which is a portable toolkit for speech recognition system. The proposed system decomposes the page into its text lines and then extracts a set of simple statistical features from small overlapped windows running through each text line. The feature vector sequence is injected to the global model for training and recognition purposes. A data corpus which includes Arabic text of more than 100 A4-size sheets typewritten in Tahoma font is used to assess the performance of the proposed system. [ABSTRACT FROM AUTHOR]

The polynomial neural network, or called polynomial network classifier (PNC), is a powerful nonlinear classifier that can separate classes of complicated distributions. A method that expands polynomial terms on principal subspace has yielded superior performance. In this paper, we aim to further improve the performance of the subspace-feature-based PNC. In the framework of discriminative feature extraction (DFE), we adjust the subspace parameters together with the network weights in supervised learning. Under the objective of minimum squared error, the parameters can be efficiently updated by stochastic gradient descent. In experiments on 13 datasets from the UCI Machine Learning Repository, we show that DFE can either improve the classification accuracy or reduce the network complexity. On seven datasets, the accuracy of PNC is competitive with support vector classifiers. [ABSTRACT FROM AUTHOR]

Object detection in real images has attracted much attention during the last decade. Using machine learning and large databases it is possible to develop detectors for visual categories that have a very high hit-rate, with low false positive rates. In this paper we investigate a general probabilistic framework for context based scene interpretation using multiple detectors. Methods for finding maximum likelihood estimates of scenes given detection results are presented. Although we have investigated how the method works for a specific case, namely for face detection, it is a general method. We show how to combine the results of a number of detectors i.e. face, eye, nose and mouth detectors. The methods have been tested using detectors trained on real images, with promising results. [ABSTRACT FROM AUTHOR]

We address the problem of fusing experts employing diverse similarity measures in LDA face space. The gradient direction measure is reviewed and experimentally compared with the normalised correlation in two different conditions, when the face images are well registered and when the registration process is performed automatically. We show that by combining the gradient direction measure and normalised correlation using a confidence based gating, the resulting decision making scheme consistently outperforms the best method. The gating is based on a novel decision confidence measure proposed in the paper. [ABSTRACT FROM AUTHOR]

In this paper we demonstrate how to recover surface shape from single images of faces using shape-from-shading when shadows are present. We make use of a statistical representation of the distribution of surface normal directions based on the equidistant azimuthal projection. This is allows us to develop a statistical model of the variations in facial shape in the surface normal domain. We show how ideas from robust statistics can be used to fit the model to facial images in which there is significant self-shadowing. The method is evaluated on both synthetic and real-world images. It is demonstrated to effectively fill-in the facial surface when more than 30% of the area is subject to self-shadowing. [ABSTRACT FROM AUTHOR]

Among recent topics studied in context of feature selection the hybrid algorithms seem to receive particular attention. In this paper we propose a new hybrid algorithm, the flexible hybrid floating sequential search algorithm, that combines both the filter and wrapper search principles. The main benefit of the proposed algorithm is its ability to deal flexibly with the quality-of-result versus computational time trade-off and to enable wrapper based feature selection in problems of higher dimensionality than before. We show that it is possible to trade significant reduction of search time for negligible decrease of the classification accuracy. Experimental results are reported on two data sets, WAVEFORM data from the UCI repository and SPEECH data from British Telecom. [ABSTRACT FROM AUTHOR]