Catálogo de publicaciones - libros

Discriminant analysis diagram (DAD) as a methodology for design of error based pattern recognition systems is presented. Recognition, i.e.verification, identification, and indexing of patterns is based on their intra-class errors when pattern classes used in training time are different than classes recognized in system exploiting time. The situation is typical for biometric identity verification. DAD is a labeled directed graph with the distinguished source node, sink node, and other nodes representing various discriminant features of recognized object. The edge label represents one of vector transformations: data centering, orthogonal projection onto linear subspace, vector component scaling, and orthogonal projection onto unit sphere. Linear subspaces are spanned over global, intra-class, and inter-class errors. A path from the source node to the sink node defines a basic discrimination scheme which is identified by sequential composition of subspace projections interleaved by scaling operations and single projection onto unit sphere. In the current state of development, DAD defines 24 different linear discriminant schemes for which there exist also nonlinear kernel forms. The proposed methodology is illustrated by a face verification system.

This paper presents an extension of linear discriminant analysis to higher order tensors that enables robust color object recognition. Given a labeled sample of training images, the basic idea is to consider a parallel factor model of a corresponding projection tensor. In contrast to other recent approaches, we do not compute a higher order singular value decomposition of the optimal projection. Instead, we directly derive a suitable approximation from the training data. Applying an alternating least squares procedure to repeated tensor contractions allows us to compute templates or binary classifiers alike. Moreover, we show how to incorporate a regularization method and the kernel trick in order to better cope with variations in the data. Experiments on face recognition from color images demonstrate that our approach performs very reliably, even if just a few examples are available for training.

This paper presents a new algorithm to extract the centreline of 2D objects. The algorithm computes the centreline from all the points of object in order to remain faithful to the structure of the shape. The idea is to cluster a data set constituted of the spatial position of each point composing the object. The centreline is derived from the set of computed clusters centres. The proposed method is accurate and robust to noisy boundaries.

A novel Bayesian classifier with smaller eigenvalues reset by threshold based on database is proposed in this paper. The threshold is used to substitute eigenvalues of scatter matrices which are smaller than the threshold to minimize the classification error rate with a given database, thus improving the performance of Bayesian classifier. Several experiments have shown its effectiveness. The error rates of both handwritten number recognition with MNIST database and Bengali handwritten digit recognition are small by using the proposed method. The steganalyszing JPEG images using this proposed classifier performs well.

A texture classification method using a binary texture metric is presented. The method consists of extracting local structures and describing their distribution by a global approach. Texture primitives are determined by a localized thresholding against the local median. The local spatial signature of the thresholded image is uniquely encoded as a scalar value, whose histogram helps characterize the overall texture. A multi resolution approach has been tried to handle variations in scale. Also, the encoding scheme facilitates a rich class of equivalent structures related by image rotation. Then, we demonstrate – using a set of classifications, that the proposed method significantly improves the capability of texture recognition and outperforms classical algorithms.

In this paper, we introduced new adaptive learning algorithms to extract linear discriminant analysis (LDA) features from multidimensional data in order to reduce the data dimension space. For this purpose, new adaptive algorithms for the computation of the square root of the inverse covariance matrix Σ are introduced. The proof for the convergence of the new adaptive algorithm is given by presenting the related cost function and discussing about its initial conditions. The new adaptive algorithms are used before an adaptive principal component analysis algorithm in order to construct an adaptive multivariate multi-class LDA algorithm. Adaptive nature of the new optimal feature extraction method makes it appropriate for on-line pattern recognition applications. Both adaptive algorithms in the proposed structure are trained simultaneously, using a stream of input data. Experimental results using synthetic and real multi-class multi-dimensional sequence of data, demonstrated the effectiveness of the new adaptive feature extraction algorithm.

In this paper we present a novel approach to 3D shape representation and matching utilizing a set of shape representations for 2D views of the object. The proposed technique capitalizes on the localization-preserving nature of the Hilbert space filling curve and the approximation capabilities of the Wavelet transform. Each 2D view of the object is represented by a concise 1D representation that can be used to search an image database for a match. The shape of the 3D image is represented by the set of 1D representations of its 2D views. Experimental results, on a subset of the Amsterdam Library of Object Images (ALOI) dataset, are provided.

We present an original approach for 2-D shapes description. Based on a multi-scale analysis of closed contours, this method deals with the differential turning angle.

The input contour is progressively low-pass filtered by decreasing the filter bandwidth. The output contour thus becomes increasingly smooth. At each iteration of the filtering we extract the essential points from the differential turning angle of the filtered contour to generate the d-TASS map.

Experimental results show that the d-TASS map is closely related to the contour and that it is rotation, translation and scale change invariant. It is also shearing and noise resistant. This function is local-oriented and appears to be particularly suitable for pattern recognition even for those patterns that have undergone occultation.

This paper introduces an approach for retrieving envelope of high-level object groupings in bi-level images with multiple objects. Motivated by studies in Gestalt theory, hierarchical clustering is used to detect the envelope and group its objects based on their spatial proximity, area, shape features and orientation. To decide the final grouping, the grouping outcomes are combined using an evidence accumulation method. The high-level boundary of the detected envelope is then extracted using morphological operations. For retrieval, the boundary of a query sketch is matched to the extracted envelopes from database images via dynamic space warping. Experiments on a set of bi-level logo images demonstrate the effectiveness of the approach.

We propose a novel efficient method that finds partial and complete matches to models for families of polygons in fields of corners extracted from images. The polygon models assign specific values of acuteness to each corner in a fixed-length sequence along the boundary. The absolute and relative lengths of sides can be either constrained or left unconstrained by the model. Candidate matches are found by using the model as a guide in linking corners previously extracted from images. Geometrical similarity is computed by comparing corner acutenesses and side lengths for candidate polygons to the model. Photometric similarity is derived by comparing directions of sides in candidate polygons to pixel gradient directions in the image. The flexibility and efficiency of our method is demonstrated by searching for families of buildings in large overhead images.