Catálogo de publicaciones - libros

In this paper, we present a new general method for image fusion based on Partial Differential Equation (PDE). We propose to combine pixel-level fusion and diffusion processes through one single powerful equation. The insertion of the relevant information contained in sources is achieved in the fused image by reversing the diffusion process. To solve the well-known instability problem of an inverse diffusion process, a regularization term is added. One of the advantages of such an original approach is to improve the quality of the results in case of noisy input images. Finally, few examples and comparisons with classical fusion models will demonstrate the efficiency of our method both on blurred and noisy images.

Hyperspectral images (HSI) are multidimensional and multicomponent data with a huge number of spectral bands providing spectral redundancy. To improve the efficiency of the classifiers the principal component analysis (PCA), referred to as , the maximum noise fraction (MNF) and more recently the independent component analysis (ICA), referred to as are the most commonly used techniques for dimensionality reduction (DR). But, in HSI and in general when dealing with multi-way data, these techniques are applied on the vectorized images, providing a two-way data. The spatial representation is lost and the spectral components are selected using only spectral information. As an alternative, in this paper, we propose to consider HSI as array data or tensor -instead of matrix- which offers multiple ways to decompose data orthogonally.We develop two news DR methods based on multilinear algebra tools which perform the DR using the for the first one and using the for the second one. We show that the result of spectral angle mapper (SAM) classification is improved by taking advantage of jointly spatial and spectral information and by performing simultaneously a dimensionality reduction on the spectral way and a projection onto a lower dimensional subspace of the two spatial ways.

This paper presents a learning-based algorithm for object tracking. During on-line learning we employ most informative and hard to classify examples, features maximizing individually the mutual information, stable object features within all past observations and features from the initial object template. The object undergoing tracking is discriminated by a boosted classifier built on regression stumps. We seek mode in the confidence map calculated by the strong classifier to sample new features. In a supplementing tracker based upon a particle filter we use a recursively updated mixture appearance model, which depicts stable structures in images seen so far, initial object appearance as well as two-frame variations. The update of slowly varying component is done using only pixels that are classified by the strong classifier as belonging to foreground. The estimates calculated by particle filter allow us to sample supplementary features for learning of the classifier. The performance of the algorithm is demonstrated on freely available test sequences. The resulting algorithm runs in real-time.

In this paper, we propose a collaborative technique for face orientation estimation in smart camera networks. The proposed spatiotemporal feature fusion analysis is based on active collaboration between the cameras in data fusion and decision making using features extracted by each camera. First, a head strip mapping method is proposed based on a Markov model and a Viterbi-like algorithm to estimate the relative angular differences to the face between the cameras. Then, given synchronized face sequences from several camera nodes, the proposed technique determines the orientation and the angular motion of the face using two features, namely the hair-face ratio and the head optical flow. These features yield an estimate of the face orientation and the angular velocity through simple analysis such as Discrete Fourier Transform (DFT) and Least Squares (LS), respectively. Spatiotemporal feature fusion is implemented via key frame detection in each camera, a forward-backward probabilistic model, and a spatiotemporal validation scheme. The key frames are obtained when a camera node detects a frontal face view and are exchanged between the cameras so that local face orientation estimates can be adjusted to maintain a high confidence level. The forward-backward probabilistic model aims to mitigate error propagation in time. Finally, a spatiotemporal validation scheme is applied for spatial outlier removal and temporal smoothing. A face view is interpolated from the mapped head strips, from which snapshots at the desired view angles can be generated. The proposed technique does not require camera locations to be known in prior, and hence is applicable to vision networks deployed casually without localization.

Independent Component Analysis (ICA) is a blind source separation method which is exploited for various applications in signal processing. In hyperspectral imagery, ICA is commonly employed for detection and segmentation purposes. But it is often thought to be unable to quantify abundances. In this paper, we propose an ICA-based method to estimate the anomaly abundances from the independent components. The first experiments on synthetic and real world hyperspectral images are very promising referring to the estimation accuracy and robustness.

In this paper we propose an unsupervised multiview image segmentation algorithm, combining multiple image cues including color, depth, and motion. First, the interested objects are extracted by computing a saliency map based on the visual attention model. By analyzing the saliency map, we automatically obtain the number of foreground objects and their bounding boxes, which are used to initialize the segmentation algorithm. Then the optimal segmentation is calculated by energy minimization under the min-cut/max-flow theory. There are two major contributions in this paper. First, we show that the performance of graph cut segmentation depends on the user interactive initialization, while our proposed method provides robust initialization instead of the random user input. In addition, we propose a novel energy function with a locally adaptive smoothness term when constructing the graphs. Experimental results demonstrate that subjectively good segmentation results are obtained.

In this paper, we develop a new wavelet domain statistical model for the removal of stationary noise in images. The new model is a combination of local linear projections onto bases of Principal Components, that perform a dimension reduction of the spatial neighbourhood, while avoiding the ”curse of dimensionality”. The models obtained after projection consist of a low dimensional Gaussian Scale Mixtures with a reduced number of parameters. The results show that this technique yields a significant improvement in denoising performance when using larger spatial windows, especially on images with highly structured patterns, like textures.

Forestry management makes great use of statistics concerning the individual trees making up a forest, but the acquisition of this information is expensive. Image processing can potentially both reduce this cost and improve the statistics. The key problem is the delineation of tree crowns in aerial images. The automatic solution of this problem requires considerable prior information to be built into the image and region models. Our previous work has focused on including shape information in the region model; in this paper we examine the image model. The aerial images involved have three bands. We study the statistics of these bands, and construct both multispectral and single band image models. We combine these with a higher-order active contour model of a ‘gas of circles’ in order to include prior shape information about the region occupied by the tree crowns in the image domain. We compare the results produced by these models on real aerial images and conclude that multiple bands improves the quality of the segmentation. The model has many other potential applications, to nano-technology, microbiology, physics, and medical imaging.

A new crossing detector is presented which also permits orientation estimation of the underlying structures. The method relies on well established tools such as the structure tensor, the double angle mapping and descriptors for second order variations. The performance of our joint crossing detector and multi-orientation estimator is relatively independent of the angular separation of the underlying unimodal structures.

To construct a very smooth nonseparable multiscaling function, we impose polynomial approximation order 2 and add new conditions on the polyphase highpass filters. We work with a dilation matrix generating quincunx lattices, and fix the index set. Other imposed conditions are orthogonal filter bank and balancing. We construct a smooth, compactly supported multiscaling function and multiwavelet, and test the system on a noisy image with good results.