Search

The main contributions of this paper are: A new method for recognition of non-deformable severe occluded objects.Local shape descriptor from the wavelet coefficients of the tangent angle signature.Most likely object retrieval based on an ensemble of Hidden Markov Models.Two new area restrictions for consistency checking between hypotheses and occlusion. Shape classification has multiple applications. In real scenes, shapes may contain severe occlusions, hardening the identification of objects. In this paper, a method for object recognition under severe occlusion is proposed. Occlusion is dealt with separating shapes into parts through high curvature points, then tangent angle signature is found for each part and continuous wavelet transform is calculated for each signature. Next, the best matching object is calculated for each part using Pearsons correlation coefficient as similarity measure between wavelets of the part and of the most probable objects in the database. For each probable class, a hidden Markov model is created in an ensemble through training with the one-class approach. For hypotheses validation, two area restrictions are set to enhance recognition performance. Experiments were conducted with 1500 test shape instances with different scenarios of object occlusion. Results showed the method not only was capable of identifying highly occluded shapes (60%80% overlapping) but also present several advantages over previous methods.

This paper proposes a unified framework for Active Constrained Clustering.The proposed method relies on a multiple kernels learning setting.Utilization of constraints is considered by clustering method in this paper.Constraints are queried based on confidence about center and boundary of clusters. This paper examines the issues of constrained clustering and active selection of clustering constraints in a unified approach. A fuzzy clustering method specially crafted to deal with non-spherical clusters and explicit pairwise constraints is proposed in this paper as a core clustering method. An active method for constraints selection is embedded into the core clustering method for querying beneficial constraints during clustering. The proposed approach has two major advantages relative to traditional methods. First, it considers the dependency of constraints effectiveness on the clustering algorithm by unifying both clustering and constraints selection in a uniform, principled framework. Second, a constraints selection method is embedded into the core clustering method based on the fact that constraints will be more useful if they are selected according to the current state of clustering. Experiments conducted on synthetic and real-world datasets show the effectiveness of the proposed method.

We propose a graph-based mode-seeking method to fit multi-structural data.The proposed method combines mode-seeking with preference analysis.The proposed method exploits the global structure of graphs by random walks.Experiments show the proposed method is superior to some other fitting methods. In this paper, we propose a novel graph-based mode-seeking fitting method to fit and segment multiple-structure data. Mode-seeking is a simple and effective data analysis technique for clustering and filtering. However, conventional mode-seeking based fitting methods are very sensitive to the proportion of good/bad hypotheses, while most of sampling techniques may generate a large proportion of bad hypotheses. In this paper, we show that the proposed graph-based mode-seeking method has significant superiority for geometric model fitting. We intrinsically combine mode seeking with preference analysis. This enables mode seeking to be beneficial for reducing the influence of bad hypotheses since bad hypotheses usually have larger residual values than good ones. In addition, the proposed method exploits the global structure of graphs by random walks to alleviate the sensitivity to unbalanced data. Experimental results on both synthetic data and real images demonstrate that the proposed method outperforms several other competing fitting methods especially for complex data.

Multi-scale PCA framework is proposed for filtering the virus images.A multi-scale CLBP descriptor is developed to extract the features.Images are classified by the SVM with polynomial kernel and the MPMC features.Experiments show that the proposed method is better than the conventional methods. Virus image classification is an important issue in clinical virology, highly accurate algorithm of automatic virus image classification is very helpful. In this paper, instead of extracting virus feature from the original image, we propose a novel method that extracts the virus feature from the filtered images by multi-scale principal component analysis (PCA). Firstly, multi-scale PCA filters are learned from all original images in the data set. Secondly, the original images are convolved with the learned filters. Therefore, the filtered images can capture the principal texture information from different perspectives. Then, the completed local binary pattern (CLBP) descriptor is firstly utilized to depict the features of all filtered virus images. The multi-scale CLBP features extracted from filtered images by multi-scale PCA are combined as the feature MPMC (Multi-scale PCA and Multi-scale CLBP), which is proposed in this paper. Finally, support vector machine (SVM) with polynomial kernel is used for classification. Experiments show that the classification accuracy based on MPMC outperforms the previous methods in the literature for the same virus image data set.

An unsupervised procedure to local symmetry detection in natural images is proposed.Achieved via a Hough-style voting approach made at different scales and coordinate spaces.A final denoising scheme aims at removing noise and spurious local symmetries.Experiments have been conducted on the recent extension of the Berkeley Segmentation Dataset.The proposed method performs well compared to the state-of-the-art algorithms. Display Omitted In this paper, we propose a new unsupervised and simple approach to local symmetry detection of ribbon-like structure in natural images. The proposed model consists in quantifying the presence of a partial medial axis segment, existing between each pair of (preliminary detected) line segments delineating the boundary of two textured regions, by a set of heuristics related both to the geometrical structure of each pair of line segments and its ability to locally delimit a homogeneous texture region in the image. This semi-local approach is finally embedded in a two-step algorithm with an amplification step, via a Hough-style voting approach achieved at different scales and coordinate spaces which aims at determining the dominant local symmetries present in the image and a final denoising step, via an averaging procedure, which aims at removing noise and spurious local symmetries. The experiments, reported in this paper and conducted on the recent extension of the Berkeley Segmentation Dataset for the local symmetry detection task, demonstrate that the proposed symmetry detector performs well compared to the best existing state-of-the-art algorithms recently proposed in the literature.

The paper presents a novel texture descriptor based on Local Binary Patterns.We address the problem of rotation invariance and feature selections.The experiments are performed on standard datasets and the results are compared with the state-of-the-art.The source code to reproduce the results is made publicly available. In this paper, we present a novel rotation-invariant and computationally efficient texture descriptor called Dominant Rotated Local Binary Pattern (DRLBP). A rotation invariance is achieved by computing the descriptor with respect to a reference in a local neighborhood. A reference is fast to compute maintaining the computational simplicity of the Local Binary Patterns (LBP). The proposed approach not only retains the complete structural information extracted by LBP, but it also captures the complimentary information by utilizing the magnitude information, thereby achieving more discriminative power. For feature selection, we learn a dictionary of the most frequently occurring patterns from the training images, and discard redundant and non-informative features. To evaluate the performance we conduct experiments on three standard texture datasets: Outex12, Outex 10 and KTH-TIPS. The performance is compared with the state-of-the-art rotation invariant texture descriptors and results show that the proposed method is superior to other approaches.

This paper introduces a kernel version of the Subspace Pursuit algorithm.The proposed method, KSP, is a new iterative method for sparse regression.KSP outperforms and is less computationally intensive than related kernel methods. Recently, results from sparse approximation theory have been considered as a means to improve the generalization performance of kernel-based machine learning algorithms. In this paper, we present Kernel Subspace Pursuit (KSP), a new method for sparse non-linear regression. KSP is a low-complexity method that iteratively approximates target functions in the least-squares sense as a linear combination of a limited number of elements selected from a kernel-based dictionary. Unlike other kernel methods, by virtue of KSP's algorithmic design, the number of KSP iterations needed to reach the final solution does not depend on the number of basis functions used nor that of elements in the dictionary. We experimentally show that, in many scenarios involving learning synthetic and real data, KSP is less complex computationally and outperforms other kernel methods that solve the same problem, namely, Kernel Matching Pursuit and Kernel Basis Pursuit.

The paper proposes a novel Multi-objective Genetic Algorithm for Data Imputation, called MOGAImp.This is the first method that applies a multi-objective approach to data imputation.MOGAImp presents a good tradeoff between the evaluation measures studied.The results confirm the MOGAImp prevalence for utilization over conflicting evaluation measures.MOGAImp codification scheme makes possible to adapt it to different application domains. A large number of techniques for data analyses have been developed in recent years, however most of them do not deal satisfactorily with a ubiquitous problem in the area: the missing data. In order to mitigate the bias imposed by this problem, several treatment methods have been proposed, highlighting the data imputation methods, which can be viewed as an optimization problem where the goal is to reduce the bias caused by the absence of information. Although most imputation methods are restricted to one type of variable whether categorical or continuous. To fill these gaps, this paper presents the multi-objective genetic algorithm for data imputation called MOGAImp, based on the NSGA-II, which is suitable for mixed-attribute datasets and takes into account information from incomplete instances and the modeling task. A set of tests for evaluating the performance of the algorithm were applied using 30 datasets with induced missing values; five classifiers divided into three classes: rule induction learning, lazy learning and approximate models; and were compared with three techniques presented in the literature. The results obtained confirm the MOGAImp outperforms some well-established missing data treatment methods. Furthermore, the proposed method proved to be flexible since it is possible to adapt it to different application domains.

We propose an unsupervised spatio-temporal image filtering method based on mean-shift.We adapt the spatial and feature range domains to handle temporal evolution.A constraint is added on the sample's evolution to select temporal neighbors.Our method outperforms standard mean-shift by adequately considering time information.Effectiveness is demonstrated on both synthetic and real brain MRI time-series data. A new spatio-temporal filtering scheme based on the mean-shift procedure, which computes unsupervised spatio-temporal filtering for univariate feature evolution, is proposed in this paper. Our main contributions are on one hand the modification of the spatial/range domains to appropriately integrate the temporal feature into the mean-shift iterative form and on the other hand the addition of a trajectory constraint in the feature space with the use of the infinity norm. Therefore, only the samples living the same life in the feature space will converge. Major assets of the standard mean-shift framework such as convergence and bandwidth parameters adjustment are preserved. In this paper, we study the relative importance of the bandwidth parameters and the efficiency of the proposed method is assessed on synthetic data and compared to the standard mean-shift framework for spatio-temporal data filtering. The obtained results have allowed us to undertake a first study on real data, which has led to encouraging results in identifying spatio-temporal evolution of multiple sclerosis lesions appearing on T2-weighted MR images.

We propose an extracting algorithm for k nearest neighbors searching.Vector inner product instead of distance calculation for distance comparison.Extracting algorithm can integrate with any other algorithm as plug-in.Two prominent algorithms and seven models are employed to experiment.Open source of proposed algorithm using dynamic memory allocation. k Nearest neighbors (kNN) searching algorithm is widely used for finding k nearest neighbors for each point in a point cloud model for noise removal and surface curvature computation. When the number of points and their density in a point cloud model increase significantly, the efficiency of a kNN searching algorithm becomes critical to various applications, thus, a better kNN approach is needed. In order to improve the efficiency of a kNN searching algorithm, in this paper, a new strategy and the corresponding algorithm are developed for reducing the amount of target points in a given data set by extracting nearest neighbors before the search begins. The nearest neighbors of a reverse nearest neighborhood are proposed to use in extracting nearest points of a query point, avoiding repetitive Euclidean distance calculation in an extracting process for saving time and memories. For any point in the model, its initial nearest neighbors can be extracted from its reverse neighborhood using an inner product of two related vectors other than direct Euclidean distance calculations and comparisons. The initial neighbors can be its full or partial set of the all nearest neighbors. If it is a partial set, the rest can be obtained by using other fast searching algorithms, which can be integrated with the proposed approach. Experimental results show that integrating extracting algorithm proposed in this paper with other excellent algorithms provides a better performance by comparing to their performances alone.

Surface and curve skeletons are important shape descriptors with applications in shape matching, simplification, retrieval, and animation. In recent years, many surface and curve skeletonization methods for 3D shapes have been proposed. However, practical comparisons of such methods against each other and against given quality criteria are quite limited in the literature. In this paper, we compare 4 surface and 6 recent curve skeletonization methods that operate on voxel shapes. We first compare the selected methods from a global perspective, using the quality criteria established by a reference paper in the field. Next, we propose a detailed comparison that refines the gained insights by highlighting small-scale differences between skeletons obtained by different methods. Keywords: Medial axes; Surface and curve skeletons; Voxel shapes

In this paper we present a new quaternion-based encryption technique for color images. In the proposed encryption method, images are written as quaternions and are rotated in a three-dimensional space around another quaternion, which is an encryption key. The encryption process uses the cipher block chaining (CBC) mode. Further, this paper shows that our encryption algorithm enables digital fingerprinting as an additional feature. This feature follows from some artifacts which occur in the decrypted image. These artifacts can be maintained at such a low level that they remain imperceptible to the human eye and unique for each key. These artifacts are considered to be users' fingerprints and can be used to identify pirates in the event of illegal redistribution of multimedia data. Identification of pirates is performed in the form of non-blind detection. The purpose of this paper is to report a new approach to digital fingerprinting and to point out the method's existing flaws which need future research. A computer-based simulation was conducted to examine the potential of our quaternion encryption scheme for the implementation of digital fingerprinting.

Extreme Support Vector Machine (ESVM) is a nonlinear robust SVM algorithm based on regularized least squares optimization for binary-class classification. In this paper, a novel algorithm for regression tasks, Extreme Support Vector Regression (ESVR), is proposed based on ESVM. Moreover, kernel ESVR is suggested as well. Experiments show that, ESVR has a better generalization than some other traditional single hidden layer feedforward neural networks, such as Extreme Learning Machine (ELM), Support Vector Regression (SVR) and Least Squares-Support Vector Regression (LS-SVR). Furthermore, ESVR has much faster learning speed than SVR and LS-SVR. Stabilities and robustnesses of these algorithms are also studied in the paper, which shows that the ESVR is more robust and stable.

The bag of visual words model (BoW) and its variants have demonstrated their effectiveness for visual applications. The BoW model first extracts local features and generates the corresponding codebook where the elements of a codebook are viewed as visual words. However, the codebook is dataset dependent and has to be generated for each image dataset. Besides, when we only have a limited number of training images, the codebook generated correspondingly may not be able to encode images well. This requires a lot of computational time and weakens the generalization power of the BoW model. To solve these problems, in this paper, we propose to undo the dataset bias by linear codebook transformation in an unsupervised manner. To represent each point in the local feature space, we need a number of linearly independent basis vectors. We view the codebook as a linear transformation of these basis vectors. In this way, we can transform the pre-learned codebooks for a new dataset using the pseudo-inverse of the transformation matrix. However, this is an under-determined problem which may lead to many solutions. Besides, not all of the visual words are equally important for the new dataset. It would be more effective if we can make some selection and choose the discriminative visual words for transformation. Specifically, the sparsity constraints and the F-norm of the transformation matrix are used in this paper. We propose an alternative optimization algorithm to jointly search for the optimal linear transformation matrixes and the encoding parameters. The proposed method needs no labeled images from either the source dataset or the target dataset. Image classification experimental results on several image datasets show the effectiveness of the proposed method.

This paper presents focusing on an object of interest in thermal infrared (IR) imagery using the morphological gradient operator. Most existing focus metrics measure the degree of sharpness on the edge of an object in the field of view, often based on the local gradient operators of pixel brightness intensity. However, such focus measures may fail to find the optimal focusing distance to the object in thermal IR images, where strong edge components of an object do not exist. In particular, when the end goal of image acquisition is object recognition, focusing on an object must retain prominent features of the object for recognition. In this paper, the performances of various focus measures are evaluated in terms of sharpness as well as recognition accuracies for face recognition in thermal IR images. Experiment results show that the morphological gradient operator outperforms conventional gradient operators in terms of autofocusing resolution metric as well as face recognition accuracy.

The paper presents a novel unsupervised contextual spectral (CSE) framework for human action and video classification. Similar to textual words, the visual word (a mid-level semantic) representation of an image or video contains a combination of synonymous words which give rise to the ambiguity of the representation. To narrow the semantic gap between visual words (mid-level semantic representation) and high-level semantics, we propose a high level representation called approximate-semantic descriptor. The experimental results show that the proposed approach for visual words disambiguation could improve the subsequent classification performance. In the paper, the approximate-semantic descriptor learning is formulated as a spectral clustering problem, such that semantically associated visual words are placed closely in low-dimensional semantic space and then clustered into one approximate-semantic descriptor. Specifically, the high level representation of human action videos is learnt by capturing the inter-video context of mid-level semantics via a non-parametric correlation measure. Experiments on four standard datasets demonstrate that our approach can achieve significantly improved results with respect to the state of the art, particularly for unconstrained environments.

Segmenting multidimensional images, in particular hyperspectral images, is still an open subject. Two are the most important issues in this field. On one hand, most methods do not preserve the multidimensional character of the signals throughout the segmentation process. They usually perform an early projection of the hyperspectral information to a two dimensional representation with the consequent loss of the large amount of spectral information these images provide. On the other hand, there is usually very little and dubious ground truth available, making it very hard to train and tune appropriate segmentation and classification strategies. This paper describes an approach to the problem of segmenting and classifying regions in multidimensional images that performs a joint two-step process. The first step is based on the application of cellular automata (CA) and their emergent behavior over the hyperspectral cube in order to produce homogeneous regions. The second step employs a more traditional SVM in order to provide labels for these regions to classify them. The use of cellular automata for segmentation in hyperspectral images is not new, but most approaches to this problem involve hand designing the rules for the automata and, in general, average out the spectral information present. The main contribution of this paper is the study of the application of evolutionary methods to produce the CA rule sets that result in the best possible segmentation properties under different circumstances without resorting to any form of projection until the information is presented to the user. In addition, we show that the evolution process we propose to obtain the rules can be carried out over RGB images and then the resulting automata can be used to process multidimensional hyperspectral images successfully, thus avoiding the problem of lack of appropriately labeled ground truth images. The procedure has been tested over synthetic and real hyperspectral images and the results are very competitive.

In this paper, we address the challenging task of simultaneous recognition of overlapping sound events from single channel audio. Conventional frame-based methods are not well suited to the problem, as each time frame contains a mixture of information from multiple sources. Missing feature masks are able to improve the recognition in such cases, but are limited by the accuracy of the mask, which is a non-trivial problem. In this paper, we propose an approach based on Local Spectrogram Features (LSFs) which represent local spectral information that is extracted from the two-dimensional region surrounding “keypoints” detected in the spectrogram. The keypoints are designed to locate the sparse, discriminative peaks in the spectrogram, such that we can model sound events through a set of representative LSF clusters and their occurrences in the spectrogram. To recognise overlapping sound events, we use a Generalised Hough Transform (GHT) voting system, which sums the information over many independent keypoints to produce onset hypotheses, that can detect any arbitrary combination of sound events in the spectrogram. Each hypothesis is then scored against the class distribution models to recognise the existence of the sound in the spectrogram. Experiments on a set of five overlapping sound events, in the presence of non-stationary background noise, demonstrate the potential of our approach.

The area under the receiver operating characteristic curve is a widely used measure of the performance of classification rules. This paper shows that when classifications are based solely on data describing individual objects to be classified, the area under the receiver operating characteristic curve is an incoherent measure of performance, in the sense that the measure itself depends on the classifier being measured. It significantly extends earlier work by showing that this incoherence is not a consequence of a cost-based interpretation of misclassifications, but is a fundamental property of the area under the curve itself. The paper also shows that if additional information, such as the class assignments of other objects, is taken into account when making a classification, then the area under the curve is a coherent measure, although in those circumstances it makes an assumption which is seldom if ever appropriate.

Kernel k-means clustering method has been proved to be effective in identifying non-isotropic and linearly inseparable clusters in the input space. However, this method is not a suitable one for large datasets because of its quadratic time complexity with respect to the size of the dataset. This paper presents a simple prototype based hybrid approach to speed-up the kernel k-means clustering method for large datasets. The proposed method works in two stages. First, the dataset is partitioned into a number of small grouplets by using the leaders clustering method which takes the size of each grouplet, called the threshold t, as an input parameter. The conventional leaders clustering method is modified such that these grouplets are formed in the kernel induced feature space, but each grouplet is represented by a pattern (called its leader) in the input space. The dataset is re-indexed according to these grouplets. Later, the kernel k-means clustering method is applied over the set of leaders to derive a partition of the leaders set. Finally, each leader is replaced by its group to get a partition of the entire dataset. The time complexity as well as space complexity of the proposed method is O(n+p^2), where p is the number of leaders. The overall running time and the quality of the clustering result depends on the threshold t and the order in which the dataset is scanned. This paper presents a study on how the input parameter t affects the overall running time and the clustering quality obtained by the proposed method. Further, both theoretically and experimentally it has been shown how the order of scanning of the dataset affects the clustering result. The proposed method is also compared with the other recent methods that are proposed to speed-up the kernel k-means clustering method. Experimental study with several real world as well as synthetic datasets shows that, for an appropriate value of t, the proposed method can significantly reduce the computation time but with a small loss in clustering quality, particularly for large datasets.

A set of objectives are partitioned into groups by means of fuzzy set theory-based clustering approaches, which ignores the hesitancy introduced by the relationship degree between two entities. The interval-based membership generalization in vague sets (VSs) is more expressive than fuzzy sets (FSs) in describing and dealing with data vagueness. In this paper, we introduce a fuzzy clustering algorithm in the context of VSs theory and fuzzy C-means clustering (FCM), i.e., Vague C-means clustering algorithm (VCM). First, the objective function of VCM and the definition of interval-based membership function are given. Then, the QPSO (quantum-behaved particle swarm optimization)-based VCM calculation is proposed. Contrastive experimental results show that the proposed scheme is more effective and more efficient than FCM and three varieties of FCM, that is, FCM-HDGA, GK-FCM and KL-FCM. Besides, the paper also discusses the influence of the VCM parameters on the clustering results.

Binary descriptors allow faster similarity computation than real-valued ones while requiring much less storage. As a result, many algorithms have recently been proposed to binarize floating-point descriptors so that they can be searched for quickly. Unfortunately, even if the similarity between vectors can be computed fast, exhaustive linear search remains impractical for truly large databases and approximate nearest neighbor (ANN) search is still required. It is therefore surprising that relatively little attention has been paid to the efficiency of ANN algorithms on binary vectors and this is the focus of this paper. We first show that binary-space Voronoi diagrams have thick boundaries, meaning that there are many points that lie at the same distance from two random points. This violates the implicit assumption made by most ANN algorithms that points can be neatly assigned to clusters centered around a set of cluster centers. As a result, state-of-the-art algorithms that can operate on binary vectors exhibit much lower performance than those that work with floating point ones. The above analysis is the first contribution of the paper. The second one is two effective ways to overcome this limitation, by appropriately randomizing either a tree-based algorithm or hashing-based one. In both cases, we show that we obtain precision/recall curves that are similar to those than can be obtained using floating point number calculation, but at much reduced computational cost.

Person re-identification is a recently introduced computer vision task that consists of recognising an individual who was previously observed over a video-surveillance camera network. Among the open problems, in this paper we focus on computational complexity. Despite its practical relevance, especially in real-time applications, this issue has been overlooked in the literature so far. In this paper, we address it by exploiting a framework we proposed in a previous work. It allows us to turn any person re-identification method, that uses multiple components and a body part subdivision model, into a dissimilarity-based one. Each individual is represented as a vector of dissimilarity values to a set of visual prototypes, that are drawn from the original non-dissimilarity representation. Experiments on two benchmark datasets provide evidence that a dissimilarity representation provides very fast re-identification methods. We also show that, even if the re-identification accuracy can be lower (especially when the number of candidates is low), the trade-off between processing time and accuracy can nevertheless be advantageous, in real-time application scenarios involving a human operator.

This paper is a response to the similarity measure and pattern recognition problem of Mitchell that was published in Pattern Recognition Letters, 2003. The purpose of this paper is threefold. First, we reviewed and revised her computation for similarity measures. Second, we proved that the similarity values for the one-norm should be larger than that for the two-norm for her pattern recognition problem. Third, we proposed a more scattered similarity measure to help researchers determine patterns. Our findings may shed light on the ongoing debate between Li and Cheng (2002) and Mitchell (2003).

One of the most usual strategies for tackling the 3D object recognition problem consists of representing the objects by their appearance. 3D recognition can therefore be converted into a 2D shape recognition matter. This paper is focused on carrying out an in depth qualitative and quantitative analysis with regard to the performance of 2D shape recognition methods when they are used to solve 3D object recognition problems. Well known shape descriptors (contour and regions) and 2D similarities measurements (deterministic and stochastic) are thus combined to evaluate a wide range of solutions. In order to quantify the efficiency of each approach we propose three parameters: Hard Recognition Rate (Hr), Weak Recognition Rate (Wr) and Ambiguous Recognition Rate (Ar). These parameters therefore open the evaluation to active recognition methods which deal with uncertainty. Up to 42 combined methods have been tested on two different experimental platforms using public database models. A detailed report of the results and a discussion, including detailed remarks and recommendations, are presented at the end of the paper.

Highlights? We show how the use of the original objects could improve the consensus results. ? Weighted association matrix extracts more information from clustering ensembles. ? Data representations and similarity measures can be summarized into a unified matrix. ? The two clustering ensemble algorithms have good performance on several datasets. ? These algorithms are able to work with numerical, categorical and mixed data. Co-association matrix has been a useful tool in many clustering ensemble techniques as a similarity measure between objects. In this paper, we introduce the weighted-association matrix, which is more expressive than the traditional co-association as a similarity measure, in the sense that it integrates information from the set of partitions in the clustering ensemble as well as from the original data of object representations. The weighted-association matrix is the core of the two main contributions of this paper: a natural extension of the well-known evidence accumulation cluster ensemble method by using the weighted-association matrix and a kernel based clustering ensemble method that uses a new data representation. These methods are compared with simple clustering algorithms as well as with other clustering ensemble algorithms on several datasets. The obtained results ratify the accuracy of the proposed algorithms.

This paper aims to solve the problem of matching images containing repetitive patterns. Although repetitive patterns widely exist in real world images, these images are difficult to be matched due to local ambiguities even if the viewpoint changes are not very large. It is still an open and challenging problem. To solve the problem, this paper proposes to match pairs of interest points and then obtain point correspondences from the matched pairs of interest points based on the low distortion constraint, which is meant that the distortions of point groups should be small across images. By matching pairs of interest points, local ambiguities induced by repetitive patterns can be reduced to some extent since information in a much larger region is used. Meanwhile, owing to our newly defined compatibility measure between one correspondence and a set of point correspondences, the obtained point correspondences are very reliable. Experiments have demonstrated the effectiveness of our method and its superiority to the existing methods.

International audience; In some real world applications, such as spectrometry, functional models achieve better predictive performances if they work on the derivatives of order m of their inputs rather than on the original functions. As a consequence, the use of derivatives is a common practice in Functional Data Analysis, despite a lack of theoretical guarantees on the asymptotically achievable performances of a derivative based model. In this paper, we show that a smoothing spline approach can be used to preprocess multivariate observations obtained by sampling functions on a discrete and finite sampling grid in a way that leads to a consistent scheme on the original infinite dimensional functional problem. This work extends (Mas and Pumo, 2009) to nonparametric approaches and incomplete knowledge. To be more precise, the paper tackles two difficulties in a nonparametric framework: the information loss due to the use of the derivatives instead of the original functions and the information loss due to the fact that the functions are observed through a discrete sampling and are thus also unperfectly known: the use of a smoothing spline based approach solves these two problems. Finally, the proposed approach is tested on two real world datasets and the approach is experimentaly proven to be a good solution in the case of noisy functional predictors.

This paper presents an empirical evaluation on the methods of reducing the dimensionality of dissimilarity spaces for optimizing dissimilarity-based classifications (DBCs). One problem of DBCs is the high dimensionality of the dissimilarity spaces. To address this problem, two kinds of solutions have been proposed in the literature: prototype selection (PS) based methods and dimension reduction (DR) based methods. Although PS-based and DR-based methods have been explored separately by many researchers, not much analysis has been done on the study of comparing the two. Therefore, this paper aims to find a suitable method for optimizing DBCs by a comparative study. Our empirical evaluation, obtained with the two approaches for an artificial and three real-life benchmark databases, demonstrates that DR-based methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA) based methods, generally improve the classification accuracies more than PS-based methods. Especially, the experimental results demonstrate that PCA is more useful for the well-represented data sets, while LDA is more helpful for the small sample size problems.

A rules induction algorithm dedicated to describe groups of genes with similar expression profiles by means of Gene Ontology terms is discussed in the paper. The presented algorithm takes into consideration information contained in the Gene Ontology graph. A huge number of created rules requires defining the rules quality and similarity measures, thus the paper presents such measures and proposes a new method of the most interesting rules selection. Features reduction method based on the rough sets theory is adopted and applied in order to reduce the number of Gene Ontology terms occurring in rules. The paper presents results of performed experiments and describes shortly the internet application RuleGO in which the proposed methods were implemented.

This paper investigates feature selection based on rough sets for dimensionality reduction in Case-Based Reasoning classifiers. In order to be useful, Case-Based Reasoning systems should be able to manage imprecise, uncertain and redundant data to retrieve the most relevant information in a potentially overwhelming quantity of data. Rough Set Theory has been shown to be an effective tool for data mining and for uncertainty management. This paper has two central contributions: (1) it develops three strategies for feature selection, and (2) it proposes several measures for estimating attribute relevance based on Rough Set Theory. Although we concentrate on Case-Based Reasoning classifiers, the proposals are general enough to be applicable to a wide range of learning algorithms. We applied these proposals on twenty data sets from the UCI repository and examined the impact of feature selection over classification performance. Our evaluation shows that all three proposals benefit the basic Case-Based Reasoning system. They also present robustness in comparison to well-known feature selection strategies.

Wavelet shrinkage is an image denoising technique based on the concept of thresholding the wavelet coefficients. The key challenge of wavelet shrinkage is to find an appropriate threshold value, which is typically controlled by the signal variance. To tackle this challenge, a new image shrinkage approach, called AntShrink, is proposed in this paper. The proposed approach exploits the intra-scale dependency of the wavelet coefficients to estimate the signal variance only using the homogeneous local neighboring coefficients. This is in contrast to that all local neighboring coefficients are used in the conventional shrinkage approaches. Furthermore, to determine the homogeneous local neighboring coefficients, the ant colony optimization (ACO) technique is used in this paper to classify the wavelet coefficients. Experimental results are provided to show that the proposed approach outperforms several image denoising approaches developed in the literature.

In statistical pattern classification, generative approaches, such as linear discriminant analysis (LDA), assume a data-generating process (DGP), whereas discriminative approaches, such as linear logistic regression (LLR), do not model the DGP. In general, a generative classifier performs better than its discriminative counterpart if the DGP is well-specified and worse than the latter if the DGP is clearly mis-specified. In view of this, this paper presents a joint discriminative-generative modelling (JoDiG) approach, by partitioning predictor variables X into two sub-vectors, namely X"G, to which a generative approach is applied, and X"D, to be treated by a discriminative approach. This partitioning of X is based on statistical tests of the assumed DGP: the variables that clearly fail the tests are grouped as X"D and the rest as X"G. Then the generative and discriminative approaches are combined in a probabilistic rather than a heuristic way. The principle of the JoDiG approach is quite generic, but for illustrative purposes numerical studies of the paper focus on a widely-used case, in which the DGP assumes a multivariate normal distribution for each class. In this case, the JoDiG approach uses LDA for X"G and LLR for X"D. Numerical experiments on real and simulated data demonstrate that the performance of this new approach to classification is similar to or better than that of its discriminative and generative counterparts, in particular when the size of the training-set is comparable to the dimension of the data.

The Mahalanobis metric was proposed by extending the Mahalanobis distance to provide a probabilistic distance for a non-normal distribution. The Mahalanobis metric equation is a nonlinear second order differential equation derived from the geometrically local isotropic independence equation, which was proposed to define normal distributions in a manifold. We explain the equations and show experimental results of calculating the Mahalanobis metric by the Newton-Raphson method. We add error to an original probability density function and calculate the Mahalanobis metric to investigate the effect on the solution of error in a probability density function. This paper is an extended version of ''numerical analysis of Mahalanobis metric in vector space'' (Track 2 IBM Best Student Paper Award in ICPR'08; Son, J., Inoue, N., Yamashita, Y., 2008. Numerical analysis of Mahalanobis metric in vector space. In: Proc. 19th Int. Conf. on Pattern Recognition (CD-ROM)).

We present a novel Gaussianized vector representation for scene images by an unsupervised approach. Each image is first encoded as an ensemble of orderless bag of features. A global Gaussian Mixture Model (GMM) learned from all images is then used to randomly distribute each feature into one Gaussian component by a multinomial trial. The posteriors of the feature on all the Gaussian components serve as the parameters of the multinomial distribution. Finally, the normalized means of the features distributed in every Gaussian component are concatenated to form a supervector, which is a compact representation for each scene image. We prove that these supervectors observe the standard normal distribution. The Gaussianized vector representation is a more generalized form of the widely used histogram representation. Our experiments on scene categorization tasks using this vector representation show significantly improved performance compared with the histogram-of-features representation. This paper is an extended version of our work that won the IBM Best Student Paper Award at the 2008 International Conference on Pattern Recognition (ICPR 2008) (Zhou et al., 2008).

Natural images have a vast amount of visual patterns distributed in a wide spectrum of subspaces of varying complexities and dimensions. Understanding the characteristics of these subspaces and their compositional structures is of fundamental importance for pattern modeling, learning and recognition. In this paper, we start with small image patches and define two types of atomic subspaces: explicit manifolds of low dimensions for structural primitives and implicit manifolds of high dimensions for stochastic textures. Then we present an information theoretical learning framework that derives common models for these manifolds through information projection, and study a manifold pursuit algorithm that clusters image patches into those atomic subspaces and ranks them according to their information gains. We further show how those atomic subspaces change over an image scaling process and how they are composed to form larger and more complex image patterns. Finally, we integrate the implicit and explicit manifolds to form a primal sketch model as a generic representation in early vision and to generate a hybrid image template representation for object category recognition in high level vision. The study of the mathematical structures in the image space sheds lights on some basic questions in human vision, such as atomic elements in visual perception, the perceptual metrics in various manifolds, and the perceptual transitions over image scales. This paper is based on the J.K. Aggarwal Prize lecture by the first author at the International Conference on Pattern Recognition, Tempa, FL. 2008.

Fractal theory provides a powerful mathematical tool for texture segmentation. However, in spite of their increasing popularity, traditional fractal features are intrinsically of less accuracy due to the difference between the idea fractal model and the fractal reality of digital images. In this paper, we incorporated the multifractal analysis method into the idea of fractal signature, and thus proposed a novel type of texture descriptor called multifractal signature, which characterizes the variation of multifractal dimensions over spatial scales. In our approach, the local multifractal dimension of each scale was calculated by using the measurement acquired at two successive scales so that the time-consuming and less accurate least square fit was avoided. Based on three popular multifractal measurements, the differential box-counting (DBC) based multifractal signature, relative DBC based multifractal signature, and morphological multifractal signature were presented in this paper. The performance of the proposed texture descriptors was evaluated for segmentation of texture mosaics by comparing to the corresponding multifractal dimensions. The experimental results demonstrated that multifractal signatures can differentiate textured images more effectively and provide more robust segmentations.

Recovering human body poses from static images is challenging without prior knowledge of pose, appearance, background and clothing. In this paper, we propose a novel model-based upper poses recovery method via effective joints detection. In our research, three observables are firstly detected: face, skin, and torso. Then the joints are properly initialized according to the observables and some heuristic configuration constraints. Finally the sample-based Markov chain Monte Carlo (MCMC) method is employed to determine the final pose. The main contributions of this paper include a robust torso detector through maximizing a posterior estimation, effective joints initialization, and two continuous likelihood functions developed for effective pose inference. Experiments on 250 real world images show that our method can accurately recover upper body poses from images with a variety of individuals, poses, backgrounds and clothing.

The work in this paper presents a non-rigid registration approach using proposed radial basis function (RBF). The RBF is based on locally constrained cosines. The proposed function is designed to overcome the weaknesses, observed in previous RBFs. The criteria to evaluate the accuracy of transformation functions are also proposed in this paper. Results of the proposed RBF are analyzed and compared with the existing RBFs, based on the proposed evaluation criteria and some other similarity measures. We demonstrate the applicability of proposed approach for registration of medical images and image warping.

A recently proposed colour based tracking algorithm has been established to track objects in real circumstances [Zivkovic, Z., Krose, B. 2004. An EM-like algorithm for color-histogram-based object tracking. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pp. 798-803]. To improve the performance of this technique in complex scenes, in this paper we propose a new algorithm for optimally adapting the ellipse outlining the objects of interest. This paper presents a Lagrangian based method to integrate a regularising component into the covariance matrix to be computed. Technically, we intend to reduce the residuals between the estimated probability distribution and the expected one. We argue that, by doing this, the shape of the ellipse can be properly adapted in the tracking stage. Experimental results show that the proposed method has favourable performance in shape adaption and object localisation.

In this paper we extend the similarity classifier to cover Dombi aggregation operators. The similarity classifier was earlier studied with the ordered weighted averaging (OWA) operator, the generalized mean, and other operators. We concentrate on the use of Dombi operators during aggregation of similarities within the classifier. Four Dombi aggregation operators applied here include: conjunctive, disjunctive, weighted conjunctive, and the product operator. From each of these operators, we form a variant of the similarity classifier. The proposed methods were tested on four real world medical datasets which include: fertility, lung cancer, Haberman’s survival, and liver disorder. The new classifiers achieved improved classification accuracies compared with earlier methods. Compared with the classifier using the generalized mean, the proposed method achieved an improvement of 17.80 % on fertility, 5.88 % on lung cancer, 0.65 % on Haberman’s survival, and 0.29 % on liver disorder datasets.

The generation of morphological traits of plants such as the leaf length, width, perimeter, area, and petiole length are fundamental features of herbarium specimens, thus providing high-quality data to investigate plant responses to ongoing climatic change and plant history evolution. However, the existing measurement methods are primarily associated with manual analysis, which is labor-intensive and inefficient. This paper proposes a deep learning-based approach, called Deep Leaf, for detecting and pixel-wise segmentation of leaves based on the improved state-of-the-art instance segmentation approach, Mask Region Convolutional Neural Network (Mask R-CNN). Deep Leaf can accurately detect each leaf in the herbarium specimen and measure the associated morphological traits. The experimental results indicate that our automated approach can segment the leaves of different families. Compared to manual measurement done by ecologist and botanist experts, the average relative error of leaf length is 4.6%, while the average relative error of leaf width is 5.7%.

This paper introduces a novel Fisher discriminant classifier based on the interest filter representation for face recognition. Our interest Fisher classifier (IFC), which is robust to illumination and facial expression variability, applies the Fisher linear discriminant (FLD) to an augmented interest feature vector derived from interest filter representation of face images. The novelty of this paper comes from our proposed interest filter: the interest operator can reveal the local activity of the images but suffer from some drawbacks and we improve the capability of the interest operator and propose a multi-orientation and multi-scale interest filter. In particular, we carry out comparative studies of different similarity measures applied to various classifiers. We also perform comparative experimental studies of various face recognition schemes, including our novel IFC method, the Eigenfaces and the Fisherfaces methods, the combination of interest operator and the Eigenfaces method, the combination of interest operator and the Fisherfaces method, the Eigenfaces on the augmented interest feature vectors and other popular subspace methods. The feasibility of the new IFC method has been successfully tested on two data sets from the FERET and AR databases. The novel IFC method achieves the highest accuracy on face recognition on both two datasets.

This paper addresses the problem of selecting features in a visual object detection setup where a detection algorithm is applied to an input image represented by a set of features. The set of features to be employed in the test stage is prepared in two training-stage steps. In the first step, a feature extraction algorithm produces a (possibly large) initial set of features. In the second step, on which this paper focuses, the initial set is reduced using a selection procedure. The proposed selection procedure is based on a novel evaluation function that measures the utility of individual features for a certain detection task. Owing to its design, the evaluation function can be seamlessly embedded into an AdaBoost selection framework. The developed selection procedure is integrated with state-of-the-art feature extraction and object detection methods. The presented system was tested on five challenging detection setups. In three of them, a fairly high detection accuracy was effected by as few as six features selected out of several hundred initial candidates.

This paper proposes a hybrid genetic algorithm for classifier ensemble selection. In this paper, two local search operations used to improve offspring prior to replacement are proposed. The operations are parameterized in order to control the computation time. Experimental results and statistical tests demonstrate the effectiveness of the proposed hybrid genetic algorithm and related local search operations.

Statistical pattern classification techniques have been successfully applied to many practical classification problems. In real-world applications, the challenge is often to cope with patterns that lead to unreliable classification decisions. These situations occur either due to unexpected patterns, i.e., patterns which occur in the regions far from the training data or due to patterns which occur in the overlap region of classes. This paper proposes a method for estimating the reliability of a classifier to cope with these situations. While existing methods for quantifying the reliability are often solely based on the class membership probability estimated on global approximations, in this paper, the reliability is quantified in terms of a confidence interval on the class membership probability. The size of the confidence interval is calculated explicitly based on the local density of training data in the neighborhood of a test pattern. A synthetic example is given to illustrate the various aspects of the proposed approach. In addition, experimental evaluation on real data sets is conducted to demonstrate the effectiveness of the proposed approach to detect unexpected patterns. The lower bound of the confidence interval is used to detect the unexpected patterns. By comparing the performance with the state-of-the-art methods, we show our approach is well-founded.

This paper presents a system for apple color grading into four classes according to standards stipulated in China. To automatically grade apple fruit color, a laboratory machine vision system was developed, which consisted of a color CCD camera equipped with an image grab device, a bi-cone roller device controlled by a stepping motor, and a lighting source. Four images, one for every rotation of 90^o, were taken from each apple. Seventeen color feature parameters (FP) were extracted from each apple in the image processing. Three hundred and eighteen ''Fuji'' apples were examined by the system, and were divided into two sets, with 200 in ''Training set'' and 118 in ''Test set''. A method called organization feature parameter (OFP), based on formulae expression trees by using genetic algorithms (GA), was used in this paper. When the initial FP could not sensitively distinguish among different classes of apples, the FP were organized into one new OFP by using genetic algorithm. By applying the step decision tree algorithm in combination with the OFP method, high grade judgment ratios were achieved in the classification of two of four apple color grades, i.e., 'Extra', and 'Reject'. However, the grade judgment ratio for 'class I' and 'class II' was relatively low. Compared with BP-ANN and SVM, the OFPs method was more accurate than BP-ANN, but a little lower than SVM for identification results.

This paper introduces entropy as a feature for 1D signals. We propose as entropy measure the ratio between the signal's perturbation (i.e. its part within minimum and maximum grey level) and the total energy of the signal. A linear transformation of 2D signals into 1D signals is also illustrated together with the results concerning natural scene, texture and medical images from a large mammograms database. The aim of this paper is to verify if the entropy variations for a closed system can be a discriminant feature to select homogenous regions.

An optimization based interactive grayscale image colorization method is presented in this paper. It is an interactive colorization method, whereas the only thing user need to do is to provide some color hints by scribbling or seed pixels. The main contribution of this paper is that the colorization method greatly reduces computation time with the same good results in image quality by quadtree decomposition based non-uniform sampling. Moreover, by introducing a new simple weighting function to represent intensity similarity in the cost function, annoying color diffusion among different regions is alleviated. Experiments show that this method gives the same good quality of colorized images as the method of Levin et al. with a fraction of the computational cost.

Planar homography (collineation) is an image-to-image mapping that could be used to pinpoint stereo correspondences, but its usage has been limited to only planar scenes. This paper describes a mechanism that generalizes the use of planar homography for establishing stereo correspondences over curved scenes. Piecewise-linear approximation is used to describe curved scene, so that stereo correspondences over images of the scene are captured by a collection of local homographies. Unlike the classical approaches, the mechanism does not employ the smoothness heuristic in establishing correspondences, but is instead based upon the interplay of two processes in an iterative manner: (1) partitioning of the scene into local planar patches based upon the most current set of confirmed stereo correspondences; and (2) prediction of new stereo correspondences by the use of the local homographies defined by the partitions, plus confirmation of such predictions by the image data, thereby enlarging the set of confirmed correspondences in every iteration until no more improvement could be obtained. A key step of the mechanism is to decide what local planar homography, and thereby what planar patch, is to be used for correspondence prediction of any given unmatched feature in any given iteration of the mechanism. With knowledge of the epipolar geometry, homography could be defined by any 3 non-collinear feature correspondences. Thus for any given unmatched feature, there are choices of the aforementioned local homography as defined by any set of three non-collinear matched features in the vicinity of the feature. In this paper, we explore what criteria are to be used to decide which of such triplet sets of matched features should be used. We analyze what errors in correspondence prediction could come with a planar homography. In particular, we classify the errors into scene-related geometric error and computation-related algebraic error. We examine their effects by simulated data experiments, and propose two ways of deciding which local homography to adopt for any given unmatched feature point. Real image experiments show that the methods could lead to promising results.