The objective of this project is to develop a new classification approach for SAR images, relying on a local Gaussian model.
In a recent work on SAR image denoising [2], we studied the use of Gaussian model (as proposed in [1]) to model a local patch in a SAR image. The Gaussian models are provided by a dictionary and the “best” Gaussian is selected for each patch. In this project we would like to investigate the potential of this Gaussian model to define a local descriptor for classification purposes. Using these descriptors and some manually labeled images, the objective of the project is to train a SVM classifier to classify SAR images (classes are urban area, vegetation, water, …).
Steps of the project:
– study of [1] and [2]
– development of a SVM classifier based on the local Gaussian models
– comparison with another classification approach
Programming
Matlab or C/C++
with available SVM libraries
Supervision
F. Tupin and S. Tabti
References
[1] « From Learning Models of Natural Image Patches to Whole Image Restoration » Daniel Zoran and Yair Weiss, ICCV 2011 (oral presentation), PDF Supp code[2]
[2] https://hal.archives-ouvertes.fr/hal-01006733 S. Tabti, C. deledalle, L. Denis, F. Tupin, « Modeling the distribution of patches with shift-invariance: application to SAR image restoration », ICIP 2014