Catálogo de publicaciones - libros

In this paper we introduce a novel concept of shape evolution based on a semi-global shape width measure. It is perceptually motivated and helps us distinguish between different shape parts of varying importance. This measure can be integrated into an adaptive function in a flexible manner and used to achieve a shape evolution which can be controlled by the relative importance of shape parts (without detecting these parts explicitly). For instance, we can start to smooth out fine details while not blurring the large shape parts, or vice versa. This shape-preserving property cannot be achieved by the popular Gaussian smoothing (evolution based on geometric heat flow) and related variants, whose behavior is controlled by the local curvature alone. Experimental results demonstrate the behavior and power of this new shape evolution scheme.

We present the GlobalLocal (GL) transformation for closed planar curves. With this new transformation we can represent shape by means of two dimensional manifolds (surfaces) embedded into the unit cube. We explore some useful properties of the transform space and we demonstrate its high expressive power. We justify the high potential of the resulting invariant shape representations in object recognition by providing experimental results using the Kimia silhouette database.

A new system for object recognition in complex natural images is here proposed. The proposed system is based on two modules: image segmentation and region categorization. Original images (,) are first regularized by using a self-adaptive implementation of the Mumford-Shah functional so that the two parameters and . controlling the smoothness and fidelity, automatically adapt to the local scale and contrast. From the regularized image (,), a piece-wise constant image (,) representing a segmentation of the original image (,) is obtained. The obtained segmentation is a collection of different regions or silhouettes which must be categorized. Categorization is based on the detection of perceptual landmarks, which are scale invariant. These landmarks and the parts between them are transformed into a symbolic representation. Shapes are mapped into symbol sequences and a database of shapes in mapped into a set of symbol sequences. Categorization is obtained by using support vector machines. The Kimia silhouettes database is used for training and complex natural images from Martin database and collection of images extracted from the web are used for testing the proposed system. The proposed system is able to recognize correctly birds, mammals and fish in several of these cluttered images.

The current generation of satellite-based sensors produces a wealth of observations. The observations are recorded in different regions of electromagnetic spectrum, such as visual, infra-red, and microwave bands. The observations by themselves provide a snapshot of an area but a more interesting problem, from mining the observations for ecological or agricultural research, is to be able to correlate observations from different time instances. However, the sheer volume of data makes such correlation a daunting task. The task may be simplified in part by correlating geographical coordinates to observation but that may lead to omission of similar conditions in different regions. This paper reports on our work on an image search engine that can efficiently extract matching image segments from a database of satellite images. This engine is based on an adaptation of (Robust Image Search Engine) that has been used successfully in querying large databases of images. Our goal in the current work, in addition to matching different image segments, is to develop an interface that supports hybrid query mechanisms including the ones based on text, geographic, and content.

Content based image retrieval (CBIR), a technique which uses visual contents to search images from the large scale image databases, is an active area of research for the past decade. It is increasingly evident that an image retrieval system has to be domain specific. In this paper, we present an algorithm for retrieving images with respect to a database consisting of engineering/computer-aided design (CAD) models. The algorithm uses the shape information in an image along with its 3D information. A linear approximation procedure that can capture the depth information using the idea of shape from shading has been used. Retrieval of objects is then done using a similarity measure that combines shape and the depth information. Plotted precision/recall curves show that this method is very effective for an engineering database.

The proposed work exploits methods and techniques for automatic characterization of images for content-based access to personal photo libraries. Several techniques, even if not reliable enough to address the general problem of content-based image retrieval, have been proven quite robust in a limited domain such as the one of personal photo album. In particular, starting from the observation that most personal photos depict a usually small number of people in a relatively small number of different contexts (e.g. Beach, Public Garden, Indoor, Nature, Snow, City, etc...) we propose the use of automatic techniques borrowed from the fields of computer vision and pattern recognition to index images based on who is present in the scene and on the context where the picture were taken.

Experiments on a personal photo collection of about a thousand images proved that relatively simple content-based techniques lead to surprisingly good results in term of easyness of user access to the data.

We investigate image retrieval using interest point descriptors. New geographic information systems such as Google Earth and Microsoft Virtual Earth are providing increased access to remote sensed imagery. Content-based access to this data would support a much richer interaction than is currently possible. Interest point descriptors have proven surprisingly effective for a range of computer vision problems. We investigate their application to performing similarity retrieval in a ground-truth dataset manually constructed from 1-m IKONOS satellite imagery. We compare results of using quantized versus full descriptors, Euclidean versus Mahalanobis distance measures, and methods for comparing the sets of descriptors associated with query and target images.

A new method for automatic construction of image transformation, Feed Forward Genetic Image Network (FFGIN), is proposed in this paper. FFGIN evolves feed forward network structured image transformation automatically. Therefore, it is possible to straightforward execution of network structured image transformation. The genotype in FFGIN is a fixed length representation and consists of string which encode the image processing filter ID and connections of each node in the network. In order to verify the effectiveness of FFGIN, we apply FFGIN to the problem of automatic construction of image transformation which is “pasta segmentation” and compare with several method. From the experimental results, it is verified that FFGIN automatically constructs image transformation. Additionally, obtained structure by FFGIN is unique, and reuses the transformed images.

Diabetic retinopathy is a common eye disease directly associated with diabetes and one of the leading causes for blindness. One of its early indicators is the presence of exudates on the retina. In this paper we present a neural network-based approach to automatically detect exudates in retina images. A sliding windowing technique is used to extract parts of the image which are then passed to the neural net to classify whether the area is part of an exudate region or not. Principal component analysis and histogram specification are used to reduce training times and complexity of the network, and to improve the classification rate. Experimental results on an image data set with known exudate locations show good performance with a sensitivity of 94.78% and a specificity of 94.29%.

Various kernel functions on graphs have been defined recently. In this article, our purpose is to assess the efficiency of a marginalized kernel for image classification using structural information. Graphs are built from image segmentations, and various types of information concerning the underlying image regions as well as the spatial relationships between them are incorporated as attributes in the graph labeling. The main contribution of this paper consists in studying the impact of fusioning kernels for different attributes on the classification decision, while proposing the use of fuzzy attributes for estimating spatial relationships.