Catálogo de publicaciones - libros

We revisit the problem of model-based object recognition for intensity images and attempt to address some of the shortcomings of existing Bayesian methods, such as unsuitable priors and the treatment of residuals with a non-robust error norm. We do so by using a reformulation of the Huber metric and carefully chosen prior distributions. Our proposed method is invariant to 2-dimensional affine transformations and, because it is relatively easy to train and use, it is suited for general object matching problems.

In this paper, a multiscale representation and retrieval method for 2D shapes is introduced. First, the shapes are represented using the area of the triangles formed by the shape boundary points. Then, the Wavelet Transform (WT) is used for smoothing and decomposing the shape boundaries into multiscale levels. At each scale level, a triangle-area representation (TAR) image and the corresponding Maxima-Minima lines are obtained. The resulting multiscale TAR (MTAR) is more robust to noise, less complex, and more selective than similar methods such as the curvature scale-space (CSS). The proposed method is tested and compared to the CSS method using the MPEG-7 CE-shape-1 dataset. The results show that the proposed MTAR outperforms the CSS method for the retrieval test.

To enhance the reliability of path planning in scenery guidance system, it’s very important to select reliable or high matching probability areas from the navigation reference images for performing unmanned aerial vehicles localization. This paper applies three measures and proposes a new selection scheme base on a simplified Mumford-Shah model. The proposed method artfully avoids selecting thresholds to separate the feature images and optimally selects robust-matching areas by evolving the level set function. Experiments of the selection show that the proposed method is efficient.

In this study, we propose two algorithms for measuring the distance between shape boundaries. In the algorithms, shape boundary is represented by the Beam Angle Statistics (BAS), which maps 2-D shape information into a set of 1-D functions. Firstly, we adopt Dynamic Time Warping method to develop an efficient distance calculation scheme, which is consistent with the human visual system in perceiving shape similarity. Since the starting point of the representations may differ in shapes, the best correspondence of items is found by shifting one of the feature vectors. Secondly, we propose an approximate solution, which utilizes the cyclic nature of the shape boundary and eliminates the shifting operation. The proposed method measures the distance between the features approximately and decreases the time complexity substantially. The experiments performed on MPEG-7 Shape database show that both algorithms using BAS features outperform all the available methods in the literature.

In image deformation, one of the challenges is to produce a deformation that preserves image topology. Such deformations are called “homeomorphic”. One method of producing homeomorphic deformations is to move the pixels according to a continuous velocity field defined over the image. The pixels flow along solution curves. Finding the pixel trajectories requires solving a system of differential equations (DEs). Until now, the only known way to accomplish this is to solve the system approximately using numerical time-stepping schemes. However, inaccuracies in the numerical solution can still result in non-homeomorphic deformations. This paper introduces a method of solving the system of DEs exactly over a triangular partition of the image. The results show that the exact method produces homeomorphic deformations in scenarios where the numerical methods fail.

Image classification is often used to extract information from multi-spectral satellite images. Unsupervised methods can produce results well adjusted to the data, but that are usually difficult to assess. The purpose of this work was to evaluate the Xu internal similarity index ability to estimate the natural number of classes in multi-spectral satellite images. The performance of the index was initially tested with data produced synthetically. Four Landsat TM image sections were then used to evaluate the index. The test images were classified into a large number of classes, using the unsupervised algorithm ISODATA, which were subsequently structured hierarchically. The Xu index was used to identify the optimum partition for each test image. The results were analysed in the context of the land cover types expected for each location.

Generally, the performance of present day computer vision systems is still very much affected by varying brightness and light source conditions. Recently, Koenderink suggested that this weakness is due to methodical flaws in low level image processing. As a remedy, he develops a new theory of image modeling. This paper reports on applying his ideas to the problem of illumination insensitive face detection. Experimental results will underline that even a simple and conventional method like principal component analysis can accomplish robust and reliable face detection in the presence of illumination variation if applied to curvature features computed in Koenderink’s image space.

The goal of the skeletal shape extraction algorithm presented in this paper was to obtain a concise and robust description of planar shapes for object recognition and subsequent region segmentation. The solution of this problem is proposed in the form of a piecewise-linear skeletal representation of planar shapes, which is a very economical shape description, resistant to distortions and intensity changes. A vertex growing procedure – similar to that of pixel-by-pixel region growing – have been developed to obtain rapidly piecewise linear skeletons of gray-scale object regions without their segmentation. Simultaneously, the complete planar shape of the objects of interest is extracted by a locally-adaptive binarization performed locally at the skeleton vertex areas. The vertex extraction is implemented using a visual attention operator, which can measure the saliency level of image fragments and select vertices at the local maxima of this operator.

We have developed a generic ontology of objects, and a knowledge base of everyday physical objects. Objects are represented as assemblies of functional features and their spatial relations. Generic shape information of objects and features is stored using a partial boundary representation. Form-function reasoning is applied to deduce geometric shape elements from a feature’s functions. We have also developed a generic geometric shape based object recognition method which uses many local features. The proposed recognition method considers the concept of ontology for representation of generic functions of objects. And the use of a general shape-function reasoning with context understanding enhances the performance of object recognition.

In this paper an appearance-based statistical approach for localization and classification of 3-D objects in 2-D color images with real heterogeneous backgrounds is presented. The object feature extraction is done separately for the red, green, and blue channel. We compute six dimensional local feature vectors directly from pixel values in the images using wavelet multiresolution analysis. The first and second component of the feature vectors depend on the pixel values in the red channel, the third and fourth in the green channel, and fifth and sixth in the blue channel. Then we define an object area as a function of 3-D transformations and represent the feature vectors as probability density functions. In the recognition phase we use an algorithm based on maximum likelihood estimation for object localization and classification. Experiments made on a real data set with 39600 images compare the recognition rates for the new algorithm, which uses the color information of objects, with the results in the case of gray level images.