25-229 Registration of satellite images by spatial ...

25-229 Registration of Satellite Images by Spatial Contextualisation Physical principles and image quality Mission CNES is currently studying several image processing tools for satellite missions. A key step in these missions is the accuracy of the instrument and the geo-referencing of the images. This step allows several deformations to be corrected, such as spectral imaging system shifts or satellite vibrations, especially for very large images. Corrections can generally be performed by image registration between a source image and a reference image. CNES is currently maintaining a tool for matching satellite images referred to as Pandora2D. Pandora2D can estimate image differences at the pixel level by computing the similarity between a pixel of a source image and several pixels of the target image in two directions. Different scores in two directions are computed yielding a 4D volume for all pixels of the source image. The selected displacement between the source and target images is the most important score, which is called a "Winner-Takes-All" strategy. However, computing a 4D volume offers the possibility of defining several regularizations accounting for the neighbors of each pixel. Furthermore, the computational complexity of the resulting algorithm requires restrictions on the considered neighborhood. Image registration has already been investigated for epipolar images used in photogrammetry involving solving a 3D problem. Pandora2D is also used in the mission TRISHNA whose objective is to evaluate the water resources and needs of vegetation. Finally, Pandora2D can also be used to analyze glacier displacements or landslides. From a methodological point of view, several research tracks have already been identified to solve the image registration problem raised by this PhD thesis: Registration can be performed by defining an appropriate similarity measure (such as mutual information) between two images, accounting for the statistical properties of these images. Graph theory has been used successfully for image registration since it allows the correlations between neighboring pixels of the image to be accounted. Optimal transport is receiving increasing interest in the image processing community and can be used to build similarity measures between probability distributions. Deep learning can also be used for image registration, notably with convolutional neural networks. Application Process For more information about the topics and the co-financial partner, prepare a resume, a recent transcript, and a reference letter from your M2 supervisor/engineering school director. You will be ready to apply online before March 14th, 2025, Midnight Paris time! Profile Image Processing, Signal Processing, Mathematics, Computer Science #J-18808-Ljbffr

Created: 2025-02-08