Catálogo de publicaciones - libros

This paper presents a new multi-way filtering method for multi-way images impaired by additive white noise. Instead of matrices or vectors, multidimensional images are considered as multi-way arrays also called tensors. Some noise removal techniques consist in vectorizing or matricizing multi-way data. That could lead to the loss of inter-bands relations. The presented filtering method consider multidimensional data as whole entities. Such a method is based on multilinear algebra. We adapt multi-way Wiener filtering to multidimensional images. Therefore, we introduce specific directions for tensor flattening. To this end, we extend the SLIDE algorithm to retrieve main directions of tensors, which are modeled as straight lines. To keep the local characteristics of images, we propose to adapt quadtree decomposition to tensors. Experiments on color images and on HYDICE hyperspectral images are presented to show the importance of flattening directions for noise removal in color images and hyperspectral images.

We propose a scanning procedure for fast image acquisition, based on low-rank image representations. An initial image is predicted from a low resolution scan and a smooth interpolation of the singular triplets. This is followed by an adaptive cross correlation scan, following the maximum error in the difference image. Our approach aims at reducing the scanning time for image acquisition devices that are in the single-pixel camera category. We exemplify with results from our experimental microwave, mm-wave and terahertz imaging systems.

Many filtering methods for Gaussian noise smoothing in colour images have been proposed. The common objective of these methods is to smooth out the noise while preserving the edges and details of the image. However, it can be observed that these methods, in their effort to preserve the image structures, also generate artefacts in homogeneous regions that are actually due to noise. So, these methods can perform well in image edges and details but sometimes they do not achieve the desired smoothing in homogeneous regions. In this paper we propose a method to overcome this problem. We use fuzzy concepts to build a soft-switching technique between two Gaussian noise filters: (i) a filter able to smooth out the noise near edges and fine features while properly preserving those details and (ii) a filter able to achieve the desired smoothing in homogeneous regions. Experimental results are provided to show the performance achieved by the proposed solution.

Hexagonal image structure is a relatively new and powerful approach to intelligent vision system. The geometrical arrangement of pixels in this structure can be described as a collection of hexagonal pixels. However, all the existing hardware for capturing image and for displaying image are produced based on rectangular architecture. Therefore, it becomes important to find a proper software approach to mimic hexagonal structure so that images represented on the traditional square structure can be smoothly converted from or to the images on hexagonal structure. For accurate image processing, it is critical to best maintain the image resolution after image conversion. In this paper, we present various algorithms for image conversion between the two image structures. The performance of these algorithms will be compared though experimental results.

In this article, a new approach is presented for action recognition with only one non-calibrated camera. Invariance to view point is obtained with several acquisitions of the same action. The originality of the presented approach consists of characterizing sequences by a temporal succession of semi-global features, which are extracted from “space-time micro-volumes”. The advantages of the proposed approach is the use of robust features (estimated on several frames) associated to the ability to manage actions with variable duration and to easily segment the sequences with algorithms that are specific to time varying data. For the recognition, each view of each action is modeled by an Hidden Markov Model system. Results presented on 1614 sequences of everyday life actions like “walking”, “sitting down”, “bending down”, performed by several persons validate the proposed approach.

We present a patch-based algorithm for the purpose of object classification in video surveillance. Within detected regions-of-interest (ROIs) of moving objects in the scene, a feature vector is calculated based on template matching of a large set of image patches. Instead of matching direct image pixels, we use Gabor-filtered versions of the input image at several scales. This approach has been adopted from recent experiments in generic object-recognition tasks. We present results for a new typical video surveillance dataset containing over 9,000 object images. Furthermore, we compare our system performance with another existing smaller surveillance dataset. We have found that with 50 training samples or higher, our detection rate is on the average above 95%. Because of the inherent scalability of the algorithm, an embedded system implementation is well within reach.

This paper introduces a methodology for detecting human faces with minimum constraints on the properties of the photograph and appearance of faces. The proposed method uses average face model to save the computation time required for training process. The average face is decomposed into row and column sub-matrices and then presented to the neural network. To reduce the time required for scanning the images at places where the probability of face is very low, a pre-scan algorithm is applied. The algorithm searches the faces in the image at different scales for detecting faces in different sizes. Arbitration between multiple scales and heuristics improves the accuracy of the algorithm. Experimental results are presented in this paper to illustrate the performance of the algorithm including accuracy and speed in detecting faces.

Multi-camera networks bring in potentials for a variety of vision-based applications through provisioning of rich visual information. In this paper a method of image segmentation for human gesture analysis in multi-camera networks is presented. Aiming to employ manifold sources of visual information provided by the network, an opportunistic fusion framework is described and incorporated in the proposed method for gesture analysis. A 3D human body model is employed as the converging point of spatiotemporal and feature fusion. It maintains both geometric parameters of the human posture and the adaptively learned appearance attributes, all of which are updated from the three dimensions of space, time and features of the opportunistic fusion. In sufficient confidence levels parameters of the 3D human body model are again used as feedback to aid subsequent vision analysis. The 3D human body model also serves as an intermediate level for gesture interpretation in different applications.

The image segmentation method described in this paper is part of the gesture analysis problem. It aims to reduce raw visual data in a single camera to concise descriptions for more efficient communication between cameras. Color distribution registered in the model is used to initialize segmentation. Perceptually Organized Expectation Maximization (POEM) is then applied to refine color segments with observations from a single camera. Finally ellipse fitting is used to parameterize segments. Experimental results for segmentation are illustrated. Some examples for skeleton fitting based on the elliptical segments will also be shown to demonstrate motivation and capability of the model-based segmentation approach for multi-view human gesture analysis.

A video-based face pose recognition framework for partially occluded faces is presented. Each pose of a person’s face is approximated using a connected low-dimensional appearance manifolds and face pose is estimated by computing the minimal probabilistic distance from the partially occluded face to sub-pose manifold using a weighted mask. To deal with partially occluded faces, we detect the occluded pixels in the current frame and then put lower weights on these occluded pixels by computing minimal probabilistic distance between given occluded face pose and face appearance manifold. The proposed method was evaluated under several situations and promising results are obtained.

In this paper we propose a novel method for head tracking in large range using a scale invariant view-based appearance model. The proposed model is populated online, and it can select key frames while the head undergoes different motions in camera-near field. We propose a robust head detection algorithm to obtain accurate head region, which is used as the view of head, in each intensity image. When the head moves far from camera, the view of head is obtained through the proposed algorithm first, and then a key frame whose view of head is most similar to that of current frame is selected to recover the head pose of current frame by coordinate adjustment. In order to improve the efficiency of the tracking method, a searching algorithm is also proposed to select key frame. The proposed method was evaluated with a stereo camera and observed a robust pose recovery when the head has large motion, even when the movement along the Z axis was about 150 cm.