In this paper, we introduce a reconstruction framework that explicitly accounts for image geometry when defining the spatial interaction between pixels in the filtering process. To this end, image structure is captured using local co-occurrence statistics and is incorporated to the enhancement algorithm in a sequential fashion using the particle filtering technique. In this context, the reconstruction process is modeled using a dynamical system with multiple states and its evolution is guided by the prior density describing the image structure. Towards optimal exploration of the image geometry, an evaluation process of the state of the system is performed at each iteration. The resulting framework explores optimally spatial dependencies between image content towards variable bandwidth image reconstruction. Promising results using additive noise models demonstrate the potentials of such an explicit modeling of the geometry.

This paper introduces a variational formulation for image denoising based on a quadratic function over kernels of variable bandwidth. These kernels are scale adaptive and reflect spatial and photometric similarities between pixels. The bandwidth of the kernels is observation-dependent towards improving the accuracy of the reconstruction process and is constrained to be locally smooth. We analyze the evolution of the noise model form the RAW space to the RGB one, by propagating it over the image formation process. The experimental results demonstrate that the use of a variable bandwidth approach and an image intensity dependent noise variance ensures better restoration quality.

This paper introduces a new denoising technique that consists in recovering the image using a filtering function adapted to the image content. The definition of such a function relies on the computation of similarity between pixels of a given neighborhood. Our contribution consists in the definition of a new similarity criterion which is more robust to noise. This measure is computed from a dictionary that is adapted to image content. The projection of the image content to this subspace are used then to define a metric between a pixel and the neighborhood ones. Very promising experimental results show the potential of our approach.

We define color sensitivity or effective color depth based on the "number of reliably distinguished colors", using ideas from information theory. This figure of merit allows the comparison of different sensors or cameras and we indicate how it can be used both for the design of imaging devices and to optimize their adaptation to the scene.

Noise is an important factor in image quality. We analyze it in images produced by digital cameras. We show that, beyond the usual standard deviation measurement, spatial correlations also convey interesting information which allows to (i) better describe the perception of the noise, (ii) analyze an unknown imaging chain. Indeed, knowledge of these spatial correlations is necessary to predict the noise after the rescaling and sampling involved in a realistic imaging chains.

We investigate the natural requirement that an automated contrast enhancement algorithm be idempotent. We show that in a restricted but natural setting, it leads to a binding of the contrast changes with the decision procedure. We obtain: a natural decision procedure for the standard gamma-transformations; a class of contrast changes generalizing these; a procedure to "learn" an algorithm from a set of good images.

We propose a new model for zooming digital images. This model, driven by a partial differential equation, performs a balance between a linear zoom on homogenous zones and an anisotropic diffusion based. zoom near edges. This allows the advantages of linear zoom models and of some nonlinear ones to be combined while leaving out their drawbacks.

We propose and evaluate a method to determine whether a given digital image is the result of a digital zoom in or more generally of a linear interpolation. This information is important to some post-processing. It also sheds some light on the actual resolution of some digital cameras.

After discussing usual approaches to measuring blur, we show theoretically that there is essentially a unique way to quantify blur by a single number and we confirm the usefulness of that measure by some experiment on a natural image.

The article describes an industrial application of a vision based 3D scene reconstruction process. The images come from independent cameras whose locations are unknown and which are placed indifferently in two different media (i.e. underwater or in the air). This process is currently used in the Poseidon system (see http://www.poseidon-tech.com) to realize an automatic CAD model of a swimming pool. Poseidon is the first computer system in the world to help in the prevention of drownings in public swimming pools. We have evaluated the reconstruction accuracy based on synthetic scenes as well as real ones. In the case of real data, results are compared with measurements made by a surveyor (with a telemeter) and demonstrate high accuracy in 3D localization and reconstruction. Lastly, it proves that computer vision modeling can be efficiently used for real time applications.

We present an algorithm that extracts the largest shape within a specific class, starting from a set of image edgels. The algorithm inherits the best-first segmentation approach. However, instead of being applicable only to shapes defined within a given class of curves, we have extended our approach to tackle more general-and complex-shapes. For example, we can now process shapes obtained from sets defined over different kinds of curves and related to one another by estimated parameters. Therefore, we go from a segmentation problem to a recognition problem. In order to reduce the complexity of the searching algorithm, we work with a linearly parameterized class of shapes. This allows us, first, to use a recursive least-squares fitting, second, to cast the problem as the search of a largest edgel subset in a directed acyclic graph, and, third, to easily introduce a priori information on the location of the edgels of the searched subset. This leads us to propose a unified approach where recognition and tracking are combined. Experiments on recognizing and tracking both left and right road boundaries demonstrate that real-time processing is achievable.

This paper sets out a new representation of an image which is contrast independent. The image is decomposed into a tree of "shapes" based on connected components of level sets, which provides a full and nonredundant representation of the image. A fast algorithm to compute the tree, the fast level lines transform (FLLT), is explained in detail. Some simple and direct applications of this representation are shown.

We present an algorithm that extracts curves from a set of edgels within a specific class in a decreasing order of their "length". The algorithm inherits the perceptual grouping approaches. But, instead of using only local cues, a global constraint is imposed on each extracted subset of edgels, that the underlying curve belongs to a specific class. In order to reduce the complexity of the solution, we work with a linearly parameterized class of curves, a function of one image coordinate. This first allows one to use recursive Kalman based fitting and, second, to cast the problem as an optimal path search in a directed graph. Experiments on finding lane-markings on roads demonstrate that real-time processing is achievable.

We present a method for approximating optical flow (velocity field) by a scaled piece-wise regular vector field. The method can be formulated in a variational manner that has a strong analogy to the Mumford and Shah (1985) functional for segmenting gray level image and in particular to the region merging minimization procedure described by Morel and Soluminini (1995). The proposed functional balances a complexity term (which can be the usual sum of the boundary length) to the "ability" of pixels to move together (that is with the same model of velocity regardless of what it is). A key point that allows the method to overcome, to some extent, the non-uniqueness of the optical flow is that the segmentation is performed without any attribution of velocities to pixels or regions. Finally, we show that the obtained velocity field is, in some cases, accurate enough to perform a "Frame-Fusion" process. We show some applications of this "super-resolution" for law enforcement.

We study a model of multiscale analysis (or scale-space) applied to movies. This model comes from a thorough formalization that has been done in the theory of scale-space of static image. This formulation has led to associate with each multiscale analysis a partial differential equation (PDE). We intend here to examine the case of movies, and to insist on the motion aspects. More precisely, it has been proved in that there exists a unique affine and morphological and Galilean invariant scale-space for movies, the AMG model. This model is described by a partial differential equation. In this paper,we focus on terms appearing in that equation. We show that this model provides a reliable definition of an optical multiscale acceleration. At the practical level, scale is interpreted as a way of characterizing reliable trajectories. As we prove by experiments,the AMG model is a riddle for decimating spurious trajectories due to any kinds of nonadditive impurities and noise. Simple discrete formulae are given to implement the model.

Classical definitions of the optical flow involve a conservation law, and a spatial smoothing of the velocities based on a norm of its gradient. We propose to use the divergence of the flow instead of using the gradient. So that the optical flow we obtained is the less divergent flow which satisfies the constraint defined by the conservation law. We propose also a model with occlusions, and show results of reconstructions based on the obtained estimation of the flow.