Catálogo de publicaciones - libros

In this article, a new image segmentation model using the level set framework is introduced. The new model is region based, but it uses a new kind of region representative: local linear approximations. When compared to global image segmentation models, not only our new model manages best with local information, it is also more general. Local linear approximations allow for an excellent approximation of smooth region at global scope because they are not restricted to be linear as a whole. Our new model also takes into account the source image gradient direction which can be a good cue toward the desired segmentation. Meanwhile, our model stays simple and relatively fast to compute. The proposed technique has been applied with promising results to both synthetic and real images.

This paper presents a novel approach to bimodal texture segmentation. The proposed approach features a local binary pattern-based scheme to transform bimodal textures into bimodal gray-scale intensities, segmentable by the Lee-Seo active contour model. This process avoids the iterative calculation of active contour equation terms derived from textural feature vectors, thus reducing the associated computational overhead. The proposed approach is region-based and invariant to the initialization of the level-set function, as it converges to a stationary global minimum. It is experimentally validated on 18 composite texture images of the Brodatz album, obtaining high quality segmentation results, whereas the convergence times are up to an order of magnitude smaller than the ones reported for other active contour approaches for texture segmentation.

Face detection is a problem in image processing that has been extensively studied in the last decades. This fact is justified due to the innumerous applications this subject has in computer vision. In this paper we present a technique to detect a face in an image where several persons may be. The images were recorded from news broadcasts from different networks. First, a model based on normalized RGB color space was built to define a skin color model: skin regions were selected from several face images and extracted statistical characteristics. Second, each pixel of the image was classified in two categories: skin and not skin. Third, the face region was identified. A MATLAB function was created to detect faces of any size, achieving high detection rates.

With the development of computer science, there is an increasing demand on the object recognition in stereo images. As a binocular image pair contains larger and more complicated information than a monocular image, the stereo vision analysis has been a difficult task. Therefore how to extract the region of user’s interest is a vital step to reduce the data redundancy and improve the robustness and reliability of the analysis. The original stereo sequences used in the paper are obtained from two parallel video cameras mounted on a vehicle driving in a residential area. This paper targets the problem of data segmentation of those stereo images. It proposes a set of algorithms to separate the foreground from the complicated changing background. Experiments show that the whole process is fast and efficient in reducing the data redundancy, and improves the overall performance for the further obstacle extraction.

A variable homography is presented which can improve mosaicing across image seams where parallax (i.e., depth variations) occur. Homographies are commonly used in image mosaicing, and are ideal when the acquiring camera has been rotated around its optical center, or when the scene being mosaiced is a plane. In most cases, however, objects in the overlapping areas of adjacent images have different depths and exhibit parallax, and so a single homography will not result in a good merge.To compensate for this, an algorithm is presented which can adjust the scale of the homography relating two views, based upon scene content. The images are first rectified so that their retinal planes are parallel. The scale of the homography is then estimated for each vertical position of a sliding window over the area of overlap, and the scaled homography is applied to this region. The scale is blended so that there are no abrupt changes across the homography field, and the images are finally stitched together. The algorithm has been implemented and tested, and has been found to be effective, fast, and robust.

The problem of 3D reconstruction of nonrigid objects from uncalibrated image sequences is addressed in the paper. Our method assumes that the nonrigid object is composed of rigid part and deformation part. We first recover the affine structure of the object and separate the rigid features from the deformation ones. Then stratify the structure from affine to Euclidean space by virtue of the rigid features. The novelty of the paper lies in two aspects. First, we propose a deformation weight constraint to the problem and prove that the recovered shape bases and shapes are transformation invariant under this constraint. Second, we propose a constrained power factorization (CPF) algorithm to recover the deformation structure in affine space. The algorithm overcomes the limitations of previous SVD-based methods and can work with missing data in the tracking matrix. Extensive experiments on synthetic data and real sequences validate the proposed method and show improvements over existing solutions.

This paper describes the visual control of a parallel robot called “RoboTenis”. The system has been designed and built in order to carry out tasks in three dimensions and dynamical environments; the system is capable to interact with objects which move up to 1m/s. The control strategy is composed by two intertwined control loops: The internal loop is faster and considers the information from the joints, its sample time is 0.5 ms. Second loop represents the visual servoing system, is external to the first mentioned and represents our study purpose, it is based on predicting the object velocity which is obtained form visual information, its sample time is 8.3 ms. Lyapunov stability analysis, system delays and saturation components has been taken into account.

In this paper, we present a robust contour tracking method using a modified snake model in stereo image sequences. The main obstacle preventing typical snake-based methods from converging to boundary concavities with gourd shapes is the lack of sufficient energy near the concavities. Moreover, previous methods suffer drawbacks such as high computation cost and inefficiency with cluttered backgrounds. Our proposed method solves the problem utilizing the binormal vector and disparity information. In addition, we apply an optimization scheme on the number of snake points to better describe the object’s boundary, and we apply a region similarity energy to handle cluttered backgrounds. The proposed method can successfully define the contour of the object, and can track the contour in complex backgrounds. Performance of the proposed method has been verified with a set of experiments.

Background modeling is one of the most challenging and time consuming tasks in moving object detection for video surveillance. In this paper, we present a new algorithm which does not require any background model. Instead, it utilizes three most recent consecutive frames to detect the presence of moving object by extracting moving edges. In the proposed method, we introduce an edge segment based approach instead of traditional edge pixel based approach. We also utilize an efficient edge-matching algorithm which reduces the variation of edge localization in different frames. Finally, regions of the moving objects are extracted from previously detected moving edges by using an efficient watershed based segmentation algorithm. The proposed method is characterized through robustness against the random noise, illumination variations and quantization error and is validated with the extensive experimental results.

We present in this paper a new approach for unsupervised feature selection for non Gaussian data controlled by a finite mixture of generalized Dirichlet distributions. We model each feature by a mixture of two Beta distributions: one relevant and depends on component labels while the second distribution is uninformative for the clustering. The relevance of each feature is then quantified by the mixture weight associated to the relevant Beta distribution. Experiments in summarizing image collections have shown the merits of our approach.