Research Projects

A precise camera model is required in order to derive accurate topography from dense disparity maps computed from a stereo pair. In remote sensing, push-broom cameras are widely used, however they are far more complex than rigid cameras based on 2D sensors, commonly used in computer vision. The absolute camera calibration problem amounts to estimating the trajectory and attitude for both images, as well as some internal parameters in some cases. The major difficulty comes from the high number of parameters involved, as the attitude can vary rapidely due to oscillations; indeed, there might be an independent set of parameters every 10 to 100 rows (and images typically have 10k rows). Therefore, performing an absolute calibration (with respect to a reference ellipsoid for instance) requires a large number of ground control points (GCP). Conversely, only a few points are required when the camera model is rigid.
We propose to address this problem by locally approximating the imaging geometry by an affine model, whose parameters are inferred within a Bayesian framework, thus allowing for error estimation. We assume that a minimum number of GCPs is provided for each patch of the image where the affine model is valid. Existing elevation models (SRTM for Earth, MOLA for Mars) can be used as large sets of GCPs. The proposed technique actually performs a registration between the high resolution reconstructed elevation model and the lower resolution, existing set of GCPs, using a height field parametrization. Finally, the global camera model will be obtained by interpolating the set of local models, thus avoiding the complex task of trajectory computation by restricting to linear geometric transforms. As the number of estimated variables is much smaller than the number of data points (since there are many points for each row), the accuracy of the reconstructed topography should be significantly higher with respect to the existing model.

Disparity maps estimated using computer vision-derived algorithms usually lack quantitative error estimates. This can be a major issue when the result is used to measure reliable physical parameters, such as topography for instance. Thus, we propose to develop a new method to infer the dense disparity map from two images. We use a probabilistic approach in order to compute uncertainties as well. Within this framework, parameters are described in terms of random variables. We start by defining a generative model for both raw observed images given all model variables, including disparities. The forward model mainly consists of warping the scene using B-Splines and using a spatially adaptive, semi-parametric radiometric change map. Then we use Bayesian inference to invert and recover the a posteriori probability density function (pdf) of the disparity map.
The main contributions are: The design of an efficient model to take into account radiometric changes between images; A multigrid processing so as to speed up the optimization process; The use of raw data instead of orthorectified imagery; Efficient approximation schemes to integrate out unwanted parameters and compute uncertainties on the result. Three applications could benefit from this disparity inference method: DEM generation from a stereo pair (along or across track), automatic calibration of pushbroom cameras, and ground deformation estimation from two images at different dates.

We propose to develop new data fusion and reconstruction methods. The originality of the project lies in considering data fusion as the estimation of a single model, of arbitrary spatial and spectral resolutions. The model is to be inferred from a number of inhomogeneous observations, possibly from different sensors under various viewing conditions. It is all about reconstructing a multispectral radiometric object that best relates to the observations and integrates all the useful information contained in the initial data.
The goal is to reconstruct a multispectral reflectance map that relates to the texture and color properties of the terrain. The object provided by the fusion-reconstruction method will be a well-sampled reflectance map, possibly super-resolved regarding both spatial and spectral sampling. Pan-sharpening is a special case where the goal is to infer a color image from a high-resolution panchromatic image and lower resolution color bands.
We will start by modeling the multiband image formation from a single reflectance map. The estimation of the model parameters and related uncertainties will be performed through hierarchical Bayesian inference. This allows to integrate the physics of the Earth terrains by including available a priori knowledge. It also involves observation models describing the onboard data acquisition process (image formation and degradation from the satellites sensor to the ground reception).

We address the clustering of multidimensional polarization encoded images. The bi-dimensional coherence of the polarization information is considered. Two analysis ways are proposed: polarization contrast enhancement and a more sophisticated image processing algorithm based on a Markovian model.
The proposed algorithms are applied to two different Mueller images acquired by a fully polarimetric imaging system. Polarimetric imaging can be used effectively to characterize objects even under low radiance conditions. It is particularly useful in determining target shapes and compositions. This work investigates the usefulness of coupling polarization information with image processing techniques as opposed to pixel-based processing usually employed. Taking into account the bidimensional consistency of polarization information proves to be very efficient. Indeed, good results were obtained by using a simple polarization contrast enhancement algorithm. On the other hand, only the Hierarchical Hidden Markov Model was able to segment properly the man-made object at very low signal level.

We develop a Bayesian approach to infer the parameters of both blur and noise from remote sensing images. The modulation transfer function (MTF) of the imaging system is modeled by a parametric function with a small number of parameters. The noise is assumed to be white, additive and Gaussian. Both blur and noise processes are supposed to be stationary. To constrain this ill-posed inverse problem, the unknown scene is modeled by a scale-invariant stochastic process governed by a fractal exponent and a global energy term.
The main novelty consists of treating all parameters as random variables whose probability density function is inferred within a fully Bayesian framework. Thus, their uncertainties are provided as well as their optimal values. The chosen approach can be summarized as the computation of marginal distributions related to the useful parameters only. This requires integrating the joint pdf with respect to all the nuisance parameters, which is achieved through Laplace approximations.
In addition, we investigate several schemes for model assessment and comparison, in order to validate this approach on real images and to propose further improvements.

We have developed an unsupervised method to segment multispectral images corrupted by a correlated non-Gaussian noise. The efficiency of the proposed Markovian quadtree-based method has been illustrated on satellite image segmentation tasks with multispectral observations, in order to update nautical charts.
The proposed method relies on a hierarchical Markovian modeling and includes the estimation of all involved parameters. The parameters of the prior model are automatically calibrated while the estimation of the noise parameters is solved by identifying generalized distribution mixtures, by means of an Iterative Conditional Estimation (ICE) procedure. Generalized Gaussian (GG) distributions are purported to model various intensity distributions of the multispectral images. They are indeed well suited to a large variety of correlated multispectral data.
Our segmentation method has been successfully applied to SPOT multispectral images. Within each segmented region, a bathymetric inversion model is then estimated to recover the water depth map. Experiments on several real images have demonstrated the efficiency of the whole process and the accuracy of the obtained results has been assessed using ground truth data.

This project deals with the unsupervised segmentation of multiband satellite images. Most of these images have the particularity to be quantized on floating-point numbers with large luminance range, on different wavelengths, with missing data in several bands. These characteristics necessitate the manipulation of large amounts of accurate data on each spectral band, which is very different from the case of 8-bit integer pixels. We present some results obtained on SPOT multispectral images, by using the Marginal Posterior Mode (MPM) estimator on a quadtree structure under Markovian assumptions. The estimation of the model parameters is then addressed with Expectation-Maximization (EM)-type algorithms, allowing for an unsupervised hyperparameter estimation. The main interest of this modeling effort lies in its generality: the algorithm handles multiwavelength floating-point data in a single upward and downward scan on the quadtree. A new aspect deals with the noise statistics that are supposed to be lognormal or Generalized Gaussian for each class. Another new aspect is the in-scale-coding of the label map.