Catálogo de publicaciones - libros

We propose an automatic segmentation method for accurately identifying lung surfaces, airways, and pulmonary vessels in chest CT images. Our method consists of four steps. First, lungs and airways are extracted by inverse seeded region growing and connected component labeling. Second, pulmonary vessels are extracted from the result of first step by gray-level thresholding. Third, trachea and large airways are delineated from the lungs by three-dimensional region growing based on partitioning. Finally, accurate lung regions are obtained by subtracting the result of third step from the result of first step. The proposed method has been applied to 10 patient datasets with lung cancer or pulmonary embolism. Experimental results show that our segmentation method extracts lung surfaces, airways, and pulmonary vessels automatically and accurately.

Defect detection by ultrasonic method is limited by the pulse width. Resolution can be improved through a deconvolution process with a priori information of the pulse or by its estimation. In this paper a regularization of the Wiener filter using wavelet shrinkage is presented for the estimation of the reflectivity function. The final result shows an improved signal to noise ratio with better axial resolution.

This paper introduces a novel approach for region segmentation. In order to represent the regions, we devise and test new features based on low and high frequency wavelet coefficients which allow to capture and judge regions using changes in brightness and texture. A fusion process through statistical hypothesis testing among regions is established in order to obtain the final segmentation. The proposed local features are extracted from image data driven by global statistical information. Preliminary experiments show that the approach can segment both texturized and regions cluttered with edges, demonstrating promising results. Hypothesis testing is shown to be effective in grouping even small patches in the process.

Content description and representation are still challenging issues for the design and management of content-based image retrieval systems. This work proposes to derive content descriptors of color images by wavelet coding and indexing of the HSV (Hue, Saturation, Value) channels. An efficient scheme for this problem has to trade between being translation and rotation invariant, fast and accurate at the same time. Based on a diverse and difficult database of 1020 color images, and a strong experimental protocol we propose a method that first divides an image into 9 rectangular regions (i.e. zoning), second it applies a wavelet transformation in each of the HSV channels. A subset of the approximation and of detail coefficients of each set is then selected. A similarity measure based on histogram intersection followed by vector distance computation for the 9 regions then evaluates and ranks the closest images of the database by content. In this paper we give the details of the this new approach and show promising results upon extensive experiments performed in our lab.

We propose an automatic framework to detect and classify highlights directly from soccer videos. Sports videos are amongst the most important events for TV transmissions and journalism, however for the purpose of archiving, reuse for sports analysts and coaches, and of main interest to the audience, the considered highlights of the match should be annotated and saved separately. This procedure is done manually by many assistants watching the match from a video. In this paper we develop an automatic framework to perform such a summarization of a soccer video using object-based features. The highlights of a soccer match are defined as shots towards any of the two goal areas, i.e. plays that have already passed the midfield area. Novel algorithms are presented to perform shot classification as long distance shot and others, highlights detection based on object-based features segmentation, and highlights classification for complete summarization of the event. Experiments are reported for complete soccer matches transmitted by TV stations in Brazil, testing for different illumination (day and night), different stadium fields, teams and TV broadcasters.

This paper presents a novel efficient multiscale vessel segmentation method using the level-set framework. This technique is based on the active contour model that evolves according to the geometric measure of vessel structures. Inspired by the multiscale vessel enhancement filtering, the prior knowledge about the vessel shape is incorporated into the energy function as a region information term. In this method, a new region-based external force is combined with existing geometric snake variation models. A new speed function is designed to precisely control the curve deformation. This multiscale method is more efficient for the segmentation of vessel and line-like structures than the conventional active contour methods. Furthermore, the whole model is implemented in a level-set framework. The solution is stable and robust for various topologic changes. This method was compared with other geometric active contour models. Experimental results of human lung CT images show that this multiscale method is accurate.

Mottling is one of the most severe printing defects in modern offset printing using coated papers. It can be defined as undesired unevenness in perceived print density. In our studies, we have implemented two methods known from the literature to quantify print mottle: the standard method for prints from office equipment and the bandpass method specially designed for mottling. Our goal was to study the performance of the methods when compared to human perception. For comparisons, we used a test set of 20 grey samples which were assessed by professional and non-professional people, and the artificial methods. The results show that the bandpass method can be used to quantify mottling of grey samples with a reasonable accuracy. However, we propose a modification to the bandpass method. The enhanced bandpass method utilizes a contrast sensitivity function for the human visual system directly in the frequency domain and the function parameters are optimized based on the human assessment. This results a significant improvement in the correlation to human assessment when compared to the original bandpass method.

In this paper we resolve the problem of automatically normalize front view photos from a database that contain images of human faces with different size, angle and position. It was used a template with a standardized inter eye distance and dimensions. We are mapping all images to this template applying a geometrical transformation. It is necessary to obtain the eyes positions on image to calculate the transforms parameters. That is not a trivial problem. We use active contour to detect the human face. After that, we apply morphological filters to highlight image signal amplitude in the eyes positions. A set of criterion is applied to select a pair of point with more possibility to be the eyes. Then, a subroutine is feed with eyes coordinates to calculate and apply the geometrical transformation. Our method was applied to 500 photos and it performs very well in the 94% of all cases.

In this paper we present a new model, designated as Association Graph, to improve document representation, facilitating the ontological dimension. We explain how to generate and use this kind of graph. Also, we analyze different document similarity measures based on this representation. A classical vector space model was used to evaluate this model and measures, investigating their strengths and weaknesses. The proposed model was found to give promising results.

A scene-based method for nonuniformity correction of infrared image sequences is developed and tested. The method uses the information embedded in the scene and performs the correction in a frame by frame Kalman Filter approach. The key assumption of the method is that the uncertainty on the input infrared irradiance integrated by each detector is solved using the spatial infrared information collected from the scene. The performance of the method is tested using infrared image sequences captured by two infrared cameras.