Elena Ranguelova

Computer-assisted tree taxonomy by automated image recognition

We present an algorithm that performs image-based queries within the domain of tree taxonomy. As such, it serves as an example relevant to many other potential applications within the field of biodiversity and photo-identification. Unsupervised matching results are produced through a chain of computer vision and image processing techniques, including segmentation and automatic shape matching. The matching itself is based on a nearest neighbours search in an appropriate feature space. Finally, we briefly report on our efforts to set up a webservice to allow the general public to perform such queries online.

Analysis and synthesis of three-dimensional Gaussian Markov random fields

A method for synthesis of 3-D Gaussian Markov random fields (GMRF) is presented. Following this, a scheme for the estimation of the parameters of the model using the method of least squares (LS) and several fast techniques for segmentation of volumetric imaging are outlined. The superior performance of the 3-D analysis algorithms over 2-D processing slice by slice is shown using both synthetic textured images and real brain magnetic resonance (MR) images.

Towards an online image-based tree taxonomy

This paper reports on a first implementation of a webservice that supports image-based queries within the domain of tree taxonomy. As such, it serves as an example relevant to many other possible applications within the field of biodiversity and photo-identification. Without any human intervention matching results are produced through a chain of computer vision and image processing techniques, including segmentation and automatic shape matching. A selection of shape features is described and the architecture of the webservice is explained. Classification techniques are presented and preliminary results shown with respect to the success rate. Necessary future enhancements are discussed. Benefits are highlighted that could result from redesigning image-based expert systems as web services, open to the public at large.

Difference Field Estimation for Enhanced 3-D Texture Segmentation

The optimization problem of finding the best match for a thin-plate block of multi-texture 3-D data in a supervised framework is studied in this paper. The textures are modelled as realizations of Gaussian Markov Random Fields (GMRFs)on 3-D lattices. The classification of the central point of the data block is performed by calculating the class probability mass function (p.m.f.s) for the block given the different texture models. The KullbackLeibler measure is proposed for the minimization of the difference between the p.m.f.s distances of the The three-dimensional (3-D) segmentation of volumetric imagery poses the challenge of estimation and compensation for the existing inter-slice difference within a multi-texture 3-D data. In this paper we propose a novel method to identify the difference field by Kullback-Leibler minimization of the distance between the class probability mass functions (p.m.f.s), calculated at thin-plate 3-D blocks of data, centered at the points of interest. and fast FFT-based technique is presented for calculation of the probability density function (p.d.f.) of the data given the model. This facilitates the calculation of the classification p.m.f.s. in a supervised framework. The estimated difference field is used to enhance the performance of a computational-volume based 3-D GMRF segmentation algorithm. The performance of the overall method is illustrated with a simulation study of mosaic of synthetic 3-D textures and MRI images of human brain.

3-D methods for difference estimation in volumetric data

The estimation of the frame or slice difference in volumetric data is an important task in applications such as tracking, registration and segmentation. In this paper, we consider the problem of difference estimation in multi-texture data defined on a 3-D lattice. Each texture is modelled via a stationary Gaussian Markov random field (GMRF). Two block-matching methods are proposed for the difference field estimation. The first method uses a 3-D cross-correlation coefficient as a similarity measure. The second method is based on minimization of the Kullback-Leihler distance between the conditional class probability mass functions (p.m.f.s) of the blocks to be matched. The performance of the methods is tested on synthetic 3-D textures and on real MRI images. Difference-compensated supervised segmentation is shown to be an important application context.