Catálogo de publicaciones - libros

The objective of this work is to provide a feasible study to develop an automatic segmentation of pedicles of vertebrae on X-ray images of scoliotic patients, with the ultimate goal of the extraction of high level primitives leading to an accurate 3D spine reconstruction based on stereo-radiographic views.

Our approach relies on coarse and fine parameter free segmentation. First, active contour is performed on a probability score table built from the input pedicle sub-space yielding to a coarse shape. The prior knowledge induced from the latter shape is introduced within a level set model to refine the segmentation, resulting in a fine shape.

For validation purposes, the result obtained by the estimation of the rotation of scoliotic deformations using the resulting fine shape is compared with a gold standard obtained by manual identification by an expert. The results are promising in finding the orientation of scoliotic deformations, and hence can be used for subsequent tools for clinicians.

This paper presents an adaptive mesh generation method from a series of transversal MR images. The adaptation process is based on the construction of a metric from the gray levels of an image. The metric is constrained by four parameters which are the minimum and maximum Euclidian length of an edge, the maximum stretching of the metric and the target edge length in the metric. The initial mesh is a regular triangulation of an MR image. This initial mesh is adapted according to the metric by choosing appropriate values for the previous set of parameters. The proposed approach provides an anisotropic mesh for which the elements are clustered near the boundaries. The experimental results show that the element’s edges of the obtained mesh are aligned with the boundaries of anatomical structures identified on the MR images. Furthermore, this mesh has approximately 80% less vertices than the mesh before adaptation with vertices mainly located in the regions of interest.

Ultrasound imaging of thyroid gland provides the ability to acquire valuable information for medical diagnosis. This study presents a novel scheme for the analysis of longitudinal ultrasound images aiming at efficient and effective computer-aided detection of thyroid nodules. The proposed scheme involves two phases: a) application of a novel algorithm for the detection of the boundaries of the thyroid gland and b) detection of thyroid nodules via classification of Local Binary Pattern feature vectors extracted only from the area between the thyroid boundaries. Extensive experiments were performed on a set of B-mode thyroid ultrasound images. The results show that the proposed scheme is a faster and more accurate alternative for thyroid ultrasound image analysis than the conventional, exhaustive feature extraction and classification scheme.

In this paper we present an approach to perform automated analysis of nematodes in population images. Occlusion, shape variability and structural noise make reliable recognition of individuals a task difficult. Our approach relies on shape and geometrical statistical data obtained from samples of segmented lines. We study how shape similarity in the objects of interest, is encoded in active contour energy component values and exploit them to define shape features. Without having to build a specific model or making explicit assumptions on the interaction of overlapping objects, our results show that a considerable number of individual can be extracted even in highly cluttered regions when shape information is consistent with the patterns found in a given sample set.

Random Amplified Polymorphism DNA (RAPD) analysis is a well-known method for studying genetic relationships between individuals. In this context, processing the underlying RAPD images is a quite difficult problem, affected by various factors. A method for processing RAPD images is proposed, which aims to detect bands by pre-processing the images by performing a template correction step and a band detection mechanism. The results on RAPD images, which aim to verify genetic identity among tree individuals, show that the proposed method detects either negative or positive amplification (bands) with a sensitivity of over 94%.

A scale-adaptive method for tree segmentation and recognition based on the LiDAR height data is described. The proposed method uses an isotropic matched filtering operator optimized for the fast and reliable detection of local and multiple objects. Sequential local maxima of this operator indicate the centers of potential objects of interest such as the trees. The maxima points also represent the seed pixels for the region-growing segmentation of tree crowns. The tree verification (recognition) stage consists of tree feature estimation and comparison with reference values. Various non-uniform tree characteristics are taken into account when making decision about a tree presence in the found location. Experimental examples of the application of this method for the tree detection in LiDAR images of forests are provided.

An iterative and localized Radon transform is proposed in this paper for the specific application of road network extraction from high resolution satellite imagery. Based on an accurate estimation of the line width and line parameters in the radon space, the localized Radon transform makes it possible to detect the small road segments and the long curvilinear lines, which is a difficult task in road detection. Experiments on both synthetic and real-world imagery have shown that the proposed methodology is effective in detecting road centerlines from classified imagery.

This paper presents a novel algorithm of using wavelet transform and partial distance search (PDS) to realize the NN classifier on field programmable gate array (FPGA) with multiple modules. The algorithm identifies first closest vectors in the design set of a NN classifier for each input vector by performing the PDS in the wavelet domain, and allows concurrent classification of different input vectors for further computation acceleration by employing multiple-module PDS. For the effective reduction of the area complexity and computation latency, we proposed a novel PDS algorithm well-suited for hardware implementation and also employ subspace search, bitplane reduction and multiple-coefficient accumulation techniques. The proposed realization has been embedded in a softcore CPU for physical performance measurements. Experimental results show that the proposed realization not only provides a cost-effective solution to the FPGA implementation of NN classification systems, but also meets both high throughput and low area cost.

In this paper a solution to detect wrong way drivers on highways is presented. The proposed solution is based on three main stages: Learning, Detection and Validation. Firstly, the orientation pattern of vehicles motion flow is learned and modelled by a mixture of gaussians. The second stage (Detection and Temporal Validation) applies the learned orientation model in order to detect objects moving in the lane’s opposite direction. The third and final stage uses an Appearance-based approach to ensure the detection of a vehicle before triggering an alarm. This methodology has proven to be quite robust in terms of different weather conditions, illumination and image quality. Some experiments carried out with several movies from traffic surveillance cameras on highways show the robustness of the proposed solution.

Stroke recognition in tennis is important for building up statistics of the player and also quickly analyzing the player. It is difficult primarily on account of low resolution, variability in strokes of the same player as well as among players, variations in background, weather and illumination conditions. This paper proposes a technique to automatically classify tennis strokes efficiently under these varying circumstances. We use the geometrical information of the player to classify the strokes. The player is modeled using a color histogram and tracked across the video using histogram back projection. The binarized (segmented) output of the tracker is skeletonized and the gradient information of the skeleton is extracted to form a feature vector. A three class SVM classifier is then used to classify the stroke to be a Forehand, Backhand or Neither. We evaluated the performance of our approach with real world datasets and have obtained promising results. Finally, the proposed approach is real time and can be used with live tennis broadcasts.