Catálogo de publicaciones - libros

We propose a factorization structure from motion (SfM) framework which employs 3D active shape constraints for a 3D face model application. Two types of shape model, individual shape models and a generic model, are used to approximate non-linear manifold variation. When the 3D shape models are trained, they help the SfM algorithm to reconstruct the 3D face structure under noisy observation (tracking) circumstances. By minimizing two sets of errors, the reconstruction error generated by the linear transform of the shape models and projection error obtained by re-projecting the 3D shape to 2D positions, the 3D face shapes can be recovered optimally. Experimental results show that this algorithm accurately reconstructs the 3D shape of familiar and non-familiar faces from video sequences under circumstances of imperfect face tracking or noisy observations.

We propose an approach dedicated to recognize characters from binary images by an hybrid strategy. A statistical method is used to identify the global shape of each alphanumeric symbol. The recognition is managed by a Hierarchical Neural Network (HNN), that is able to deal with topological errors in the contour extraction. This strategy is extremely efficient for the majority of the classes: the recognition rate reaches about 99.5%. However, the performances sensitively decrease for ’similar characters’, ’8’/’B’. In that case, we adopt a strategy that revolves around decomposing the characters into structural elements. The Reeb graph generated from the binary images and a simple polygonal approximation permit to capture both topological and geometrical relevant features. The classification stage is carried out by a boosting algorithm.

In detecting the boundary of an object in an image, if certain prior shape knowledge of the object is available, an effective approach is to have the intensity gradient information in the image and the prior shape knowledge be combined together to drive an active contour for the purpose. While in the classical methods the two terms are almost always summed with a certain weight between them to form the optimization functional, in the method we propose, they are multiplied together so as to avoid the need and thus design of the weight parameter. We show that the object detection result in the traditional formulation could indeed be very much affected by the weight value, and the proposed method, being without its presence, is therefore free from the influence of the important parameter. Experimental results on cells in real biological images, whose boundaries are blurred to very different degrees across the image by the inevitably uneven illumination, are shown to demonstrate the improvement in performance.

Ears are a new biometric with major advantage in that they appear to maintain their shape with increased age. Any automatic biometric system needs enrolment to extract the target area from the background. In ear biometrics the inputs are often human head profile images. Furthermore ear biometrics is concerned with the effects of partial occlusion mostly caused by hair and earrings. We propose an ear enrolment algorithm based on finding the elliptical shape of the ear using a Hough Transform (HT) accruing tolerance to noise and occlusion. Robustness is improved further by enforcing some prior knowledge. We assess our enrolment on two face profile datasets; as well as synthetic occlusion.

This paper presents a method to determine efficient scan-patterns for spin images using robust multivariate regression. A large dataset is generated using scan-patterns with random radial scanlines through an oriented point and determining the corresponding classification performance. Eight features are chosen, which are used as predictor variables for a multivariate least trimmed squares regression algorithm, achieving an adjusted coefficient of determination of =0.80. The correlation coefficients are then used in an exemplary cost-benefit function of an exemplary application of the proposed method.

Image segmentation, one of the fundamental task of image processing, can be accurately solved using the level set framework. However, the computational time demands of the level set methods make them practically useless, especially for segmentation of large three-dimensional images. Many approximations have been introduced in recent years to speed up the computation of the level set methods. Although these algorithms provide favourable results, most of them were not properly tested against ground truth images. In this paper we present a comparison of three methods: the Sparse-Field method [1], Deng and Tsui’s algorithm [2] and Nilsson and Heyden’s algorithm [3]. Our main motivation was to compare these methods on 3D image data acquired using fluorescence microscope, but we suppose that presented results are also valid and applicable to other biomedical images like CT scans, MRI or ultrasound images. We focus on a comparison of the method accuracy, speed and ability to detect several objects located close to each other for both 2D and 3D images. Furthermore, since the input data of our experiments are artificially generated, we are able to compare obtained segmentation results with ground truth images.

The vision-based intelligent vehicle systems for environment perception have required integration of image data acquired from multiple cameras. We developed multiband camera, which can simultaneously obtain both images of visible color and near infrared. In this paper, we present a texture-based objects recognition under road environment scene using a multiband image. The new color feature is proposed to cluster meaningful regions of a multiband image and the texture segmentation is utilized in classification of texture-based objects. Experimental results show that the proposed method effectively recognizes the texture-based objects including roads, buildings, trees, and sky, as well as faces of pedestrians. In the future, by integrating the shape-based objects recognition, which includes pedestrians, cars, and bicycles with texture-based objects recognition, the proposed system can expand into a complex scene understanding system for vehicle environment perception.

Color and texture provide important visual information for real-time tracking of non-rigid and partially occluded objects. Recent developments have shown the robustness and effectiveness of color based tracking algorithms, especially for tracking tasks where object shape exhibits dramatic variability. In this article we solve the problem by tracking color distributions of background and foreground (object) points simultaneously. The key feature of our approach is careful selection of histogram resolution (or kernel radius) on each frame of a sequence.

We present a robust method to detect and locate a speaker using a joint analysis of speech sound and video image. First, the short speech sound data is analyzed to estimate the rate of spoken syllables, and a difference image is formed using the optimal frame distance derived from the rate to detect the candidates of mouth. Then, they are tracked to positively prove that one of the candidates is the mouth; the rate of mouth movements is estimated from the brightness change profiles for the first candidate and, if both the rates agree, the three brightest parts are detected in the resulting difference image as mouth and eyes. If not, the second candidate is tracked and so on. The first-order moment of the power spectrum of the brightness change profile and the lateral shifts in the tracking are also used to check whether or not they are facial parts.

This paper deals with how to correct distortions on high resolution satellite images and optimized cartographic feature extraction. It is shown, for an Ikonos satellite image that subpixel accuracy can be obtained using rational functions with only a few accurate ground control points. These control points are taken as the centres of road roundabouts, tennis courses, swimming pools and other cartographic features using least square, for greater precision in ground-image matching. The radiometric quality is also studied, in order to examine Ikonos visualization capability and consequently the potential for map updating.