Given a satellite image, we want to make a
classification (i.e. to assign a label to each pixel) coupled with a
restoration process in order to
remove noisy pixels.
This work takes place in the general framework concerning automatic
feature extraction.
Proposed method
The proposed method belongs to variational approaches. In these
approaches, the image is usually
represented by a function depending on real variables (for
instance the grey level associated to each pixel). Hypothesis :
We assume that the number and the parameters of the classes are known
(supervised classification). We also suppose that the classes
have a Gaussian distribution, so the parameters of each
class are its mean and standard deviation. The label designing a class
is set to its mean value. Herein, we are interested in the local distribution
of intensity (grey level), but other discrimant features as the texture one for instance can be considered.
The goal is to define a functional, or energy, whose minimization
leads to a piecewise constant image, the segmented regions
being homogeneous and separated by smooth boundaries (constraint of
homogeneity with anisotropic smoothing). A region will be defined as a
set of pixels belonging to the same class.
- Proposed model
- Algorithm
- Comparison with a stochastic model
- Results : synthetic image
- Results : first satellite image
- Results : second satellite image
- Conclusion