Boyan Bonev

Information Theory in Computer Vision and Pattern Recognition

Information theory has proved to be effective for solving many computer visionand pattern recognition (CVPR) problems (such as image matching, clustering and segmentation, saliency detection, feature selection, optimal classifier design and many others). Nowadays, researchers are widely bringing information theory elements to the CVPR arena. Among these elements there are measures (entropy, mutual information), principles (maximum entropy, minimax entropy) and theories (rate distortion theory, method of types). This book explores and introduces the latter elements through an incremental complexity approach at the same time where CVPR problems are formulated and the most representative algorithms are presented. Interesting connections between information theory principles when applied to different problems are highlighted, seeking a comprehensive research roadmap. The result is a novel tool both for CVPR and machine learning researchers, and contributes to a cross-fertilization of both areas.

Feature selection, mutual information, and the classification of high-dimensional patterns: Applications to image classification and microarray data analysis

We propose a novel feature selection filter for supervised learning, which relies on the efficient estimation of the mutual information between a high-dimensional set of features and the classes. We bypass the estimation of the probability density function with the aid of the entropic-graphs approximation of Rényi entropy, and the subsequent approximation of the Shannon entropy. Thus, the complexity does not depend on the number of dimensions but on the number of patterns/samples, and the curse of dimensionality is circumvented. We show that it is then possible to outperform algorithms which individually rank features, as well as a greedy algorithm based on the maximal relevance and minimal redundancy criterion. We successfully test our method both in the contexts of image classification and microarray data classification. For most of the tested data sets, we obtain better classification results than those reported in the literature.

Constellations and the unsupervised learning of graphs

In this paper, we propose a novel method for the unsupervised clustering of graphs in the context of the constellation approach to object recognition. Such method is an EM central clustering algorithm which builds prototypical graphs on the basis of fast matching with graph transformations. Our experiments, both with random graphs and in realistic situations (visual localization), show that our prototypes improve the set median graphs and also the prototypes derived from our previous incremental method. We also discuss how the method scales with a growing number of images.

Information-theoretic selection of high-dimensional spectral features for structural recognition

Pattern recognition methods often deal with samples consisting of thousands of features. Therefore, the reduction of their dimensionality becomes crucial to make the data sets tractable. Feature selection techniques remove the irrelevant and noisy features and select a subset of features which describe better the samples and produce a better classification performance. In this paper, we propose a novel feature selection method for supervised classification within an information-theoretic framework. Mutual information is exploited for measuring the statistical relation between a subset of features and the class labels of the samples. Traditionally it has been measured for ranking single features; however, in most data sets the features are not independent and their combination provides much more information about the class than the sum of their individual prediction power. We analyze the use of different estimation methods which bypass the density estimation and estimate entropy and mutual information directly from the set of samples. These methods allow us to efficiently evaluate multivariate sets of thousands of features. Within this framework we experiment with spectral graph features extracted from 3D shapes. Most of the existing graph classification techniques rely on the graph attributes. We use unattributed graphs to show what is the contribution of each spectral feature to graph classification. Apart from succeeding to classify graphs from shapes relying only on their structure, we test to what extent the set of selected spectral features are robust to perturbations of the dataset.

Self-training and co-training applied to spanish named entity recognition

The paper discusses the usage of unlabeled data for Spanish Named Entity Recognition. Two techniques have been used: self-training for detecting the entities in the text and co-training for classifying these already detected entities. We introduce a new co-training algorithm, which applies voting techniques in order to decide which unlabeled example should be added into the training set at each iteration. A proposal for improving the performance of the detected entities has been made. A brief comparative study with already existing co-training algorithms is demonstrated.

Contextual visual localization: cascaded submap classification, optimized saliency detection, and fast view matching

In this paper, we present a novel coarse-to-fine visual localization approach: contextual visual localization. This approach relies on three elements: (i) a minimal-complexity classifier for performing fast coarse localization (submap classification); (ii) an optimized saliency detector which exploits the visual statistics of the submap; and (iii) a fast view-matching algorithm which filters initial matchings with a structural criterion. The latter algorithm yields fine localization. Our experiments show that these elements have been successfully integrated for solving the global localization problem. Context, that is, the awareness of being in a particular submap, is defined by a supervised classifier tuned for a minimal set of features. Visual context is exploited both for tuning (optimizing) the saliency detection process, and to select potential matching views in the visual database, close enough to the query view.

A novel information theory method for filter feature selection

In this paper, we propose a novel filter for feature selection. Such filter relies on the estimation of the mutual information between features and classes. We bypass the estimation of the probability density function with the aid of the entropic-graphs approximation of Renyi entropy, and the subsequent approximation of the Shannon one. The complexity of such bypassing process does not depend on the number of dimensions but on the number of patterns/samples, and thus the curse of dimensionality is circumvented. We show that it is then possible to outperform a greedy algorithm based on the maximal relevance and minimal redundancy criterion. We successfully test our method both in the contexts of image classification and microarray data classification.

Modeling deformable gradient compositions for single-image super-resolution

We propose a single-image super-resolution method based on the gradient reconstruction. To predict the gradient field, we collect a dictionary of gradient patterns from an external set of images. We observe that there are patches representing singular primitive structures (e.g. a single edge), and non-singular ones (e.g. a triplet of edges). Based on the fact that singular primitive patches are more invariant to the scale change (i.e. have less ambiguity across different scales), we represent the non-singular primitives as compositions of singular ones, each of which is allowed some deformation. Both the input patches and dictionary elements are decomposed to contain only singular primitives. The compositional aspect of the model makes the gradient field more reliable. The deformable aspect makes the dictionary more expressive. As shown in our experimental results, the proposed method outperforms the state-of-the-art methods.

Information-geometric graph indexing from bags of partial node coverages

In a previous work we have uncovered some of the most informative spectral features (Commute Times, Fiedler eigenvector, Perron-Frobenius eigenvector and Node Centrality) for graph discrimination. In this paper we propose a method which exploits information geometry (manifolds and geodesics) to characterize graphlets with covariance matrices involving the latter features. Once we have the vectorized covariance matrices in the tangent space each graph is characterized by a population of vectors in such space. Then we exploit bypass informationtheoretic measures for estimating the dissimilarities between populations of vectors. We test this measure in a very challenging database (GatorBait).

Region and constellations based categorization of images with unsupervised graph learning

In this paper, we address the problem of image categorization with a fast novel method based on the unsupervised clustering of graphs in the context of both region-based segmentation and the constellation approach to object recognition. Such method is an EM central clustering algorithm which builds prototypical graphs on the basis of either Softassign or fast matching with graph transformations. We present two realistic applications and their experimental results: categorization of image segmentations and visual localization. We compare our graph prototypes with the set median graphs. Our results reveal that, on the one hand, structure extracted from images improves appearance-based visual localization accuracy. On the other hand, we show that the cost of our central graph clustering algorithm is the cost of a pairwise algorithm. We also discuss how the method scales with an increasing amount of images. In addition, we address the scientific question of what are the bounds of structural learning for categorization. Our in-depth experiments both for region-based and feature-based image categorization, will show that such bounds depend hardly on structural variability.