Catálogo de publicaciones - libros

The advances in robotics and imaging technologies have brought various imaging devices to the field of the unmanned exploration of new environments. Forward looking sonar is one of the newly emerging imaging methods employed in the exploration of underwater environments. While the video sequences produced by forward looking sonar systems are characterized by low signal-to-noise ratio, low resolution and limited range of sight, it is expected that video enhancement techniques will facilitate the interpretation of the video sequences. Since the video enhancement techniques for forward looking sonar video sequences are applicable to most of the forward looking sonar sequences, the development of such techniques is more crucial than developing techniques for optical camera video enhancement, where only specially produced video sequences can benefit the techniques. In this paper, we introduce a procedure to enhance forward looking sonar video sequences via incorporating the knowledge of the target object obtained in previously observed frames. The proposed procedure includes inter-frame registration, linearization of image intensity, and maximum a posteriori fusion of images in the video sequence. The performance of this procedure is verified by enhancing video sequences of Dual-frequency Identification Sonar (DIDSON), the market leading forward looking sonar system.

Choosing the adequate registration and simulation parameters in non-parametric image registration methods is an open question. There is no agreement about which are the optimal values (if any) for these parameters, since they depend on the images to be registered. As a result, in the literature the parameters involved in the registration process are arbitrarily fixed by the authors. The present paper is intended to address this issue. A two-step method is proposed to obtain the optimal values of these parameters, in terms of achieving in a minimum number of iterations the best trade-off between similarity of the images and smoothness of the transformation. These optimal values minimize the joint energy functional defined in a variational framework. We focus on the specific formulation of diffusion and curvature registration, but the exposed methodology can be directly applied to other non-parametric registration schemes. The proposed method is validated over different registration scenarios.

A method to calibrate camera from a single frame of planar pattern is presented in this paper. For a camera model with four intrinsic parameters and visible lens distortion, the principal point and distortion coefficients are firstly determined through analysis of the distortion in an image. The distortion is then removed. Finally, the other intrinsic and extrinsic parameters of the camera are obtained through direct linear transform followed by bundle adjustment. Theoretically, the method makes it possible to analyze the calibration result at the level of a single frame. Practically, such a method provides a easy way to calibrate a camera used in industrial vision system on line and used in desktop vision system. Experimental results of both simulated data and real images validate the method.

We propose a scanline energy minimization algorithm for stereo vision. The proposed algorithm differs from conventional energy minimization techniques in that it focuses on the relationship between local match cost solution and the energy minimization solution. The local solution is transformed into energy minimization solution through the optimization of the disparity discontinuity. In this paper, disparity discontinuities are targeted during the energy minimization instead of the disparities themselves. By eliminating and relocating the disparity discontinuities, the energy can be minimized in iterations of () where is the number of pixels. Although dynamic programming has been adequate for both speed and performance in the scan-line stereo, the proposed algorithm was shown to have better performance with comparable speed.

There have been many progresses in the stereo matching problem. However, some remaining problems still make stereo matching difficult. Occlusion is one of such problems. In this paper, we propose a new stereo matching model that addresses this problem by using an effective visibility constraint. By considering two images simultaneously, complex geometric configurations regarding the visibility of a pixel becomes simplified, so that the visibility constraint can be modeled as a pairwise MRF. Also since the proposed model enforces the consistency between two disparity maps, the final results become consistent with each other. Belief propagation is employed for the solution of the modeled pairwise MRF. Experimental results on the standard data set demonstrate the effectiveness of our approach.

Stereo correspondence is one of the basic and most important problems in computer vision. For better correspondence, we need to determine the occlusion. Recently dynamic programming on a minimal span tree (mst) structure is used to search for correspondence. We have extended this idea. First, mst is generated directly based on the color information in the image instead of converting the color image into a gray scale. Second, have treated this mst as a Bayesian Network. Novelty is attained by considering local variances of the disparity and intensity differences in the conditional Gaussians as unobserved random parameters. These parameters are iteratively inferenced by alternate estimation along the tree given a current disparity map. It is followed by dynamic programming estimation of the map given the current variance estimates thus reducing the overall occlusion. We evaluate our algorithm on the benchmark Middlebury database. The results are promising for modeling occlusion in early vision problems.

Many industrial applications involve rotations. Different from traditional measurements, we propose a novel vision method based on image motion blur to estimate angular velocity and angular displacement in this paper. First, we transform 2D rotation to 1D translational motion by sectoring rotation blurred image. Then we use mathematical models in spatial and frequency domain to analyze translation blurred images. According to mathematical models in frequency domain, there is a series of dark parallel lines on the spectrum of the translation blurred image. These dark lines are exactly related to the velocity of translation and exposure time. Furthermore, based on the geometric relationship between rotation and translation, these dark lines are also related to angular velocity. Both simulation results and real experimental results, based on the proposed method in this paper, are provided. These results demonstrate the feasibility and efficiency of proposed method.

In this paper we propose a new segment-based stereo matching algorithm using scene hierarchical structure. In particular, we highlight a previously overlooked geometric fact: the most foreground objects can be easily detected by intensity-based cost function and the farer objects can be matched using local occlusion model constructed by former recognized objects. Then the scene structure is achieved from foreground to background. Two occlusion relations are proposed to establish occlusion model and to update cost function. Image segmentation technique is adopted to increase algorithm efficiency and to decrease discontinuity of disparity map. Experiments demonstrate that the performance of our algorithm is among the state of the art stereo algorithms on various data sets.

This paper introduces a novel Gabor-based supervised locality preserving projection (GSLPP) method for face recognition. Locality preserving projection (LPP) is a recently proposed method for unsupervised linear dimensionality reduction. LPP seeks to preserve the local structure which is usually more significant than the global structure preserved by principal component analysis (PCA) and linear discriminant analysis (LDA). In this paper, we investigate its extension, called supervised locality preserving projection (SLPP), using class labels of data points to enhance its discriminant power in their mapping into a low dimensional space. The GSLPP method, which is robust to variations of illumination and facial expression, applies the SLPP to an augmented Gabor feature vector derived from the Gabor wavelet representation of face images. We performed comparative experiments of various face recognition schemes, including the proposed GSLPP method, principal component analysis (PCA) method, linear discriminant analysis (LDA) method, locality preserving projection method, the combination of Gabor and PCA method (GPCA) and the combination of Gabor and LDA method (GLDA). Experimental results on AR database and CMU PIE database show superior of the novel GSLPP method.

In fuzzy clustering, the fuzzy c-means (FCM) algorithm is the best known and most used method. Although FCM is a very useful method, it is sensitive to noise and outliers so that Wu and Yang (2002) proposed an alternative FCM (AFCM) algorithm. In this paper, we consider the AFCM algorithms with L1-norm and fuzzy covariance. These generalized AFCM algorithms can detect elliptical shapes of clusters and also robust to noise and outliers. Some numerical experiments are performed to assess the performance of the proposed algorithms. Numerical results clearly indicate the proposed algorithms to be superior to the existing methods.