BPFA algorithm

Description

Beta Process Factor Analysis algorithm. The paper could be found here.

The code is a MATLAB implementation of BPFA with a Python function to call the MATLAB code.

The Python extensions saves the MATLAB code input data, then executes the MATLAB code, then gets the output data for Python post-processing.

MATLAB code usage

Function file: BPFA/BPFA.m

Arguments

Name	Type	Description
Y	2D or 3D double array	The (possibly multi-band) image. Note that in the case of a 3D image, the last dimension is the spectrum axis.
mask	2D logical array	The sampling mask. true means the pixel has been acquired. false means it did not.
PatchSize	int	The patch width (and height).

Keyword Arguments

Name	Type	Description
iter	int	The number of iterations. Default is 100.
K	int	The number of dictionaries. Default is 128.
step	int	The step number (see note). Default is 1.

Returns

Xhat (2D or 3D double array): The reconstructed array.

Note

In the case of really big images, one has to reduce computational costs (at least when the processing exceeds 1h15m). Two methods are possible:

1. Reduce the Patch size.
2. Reduce the number of patches.

The second point is made possible by not taking all the possible patches, i.e. by reducing overlaping. The step optional parameter set the number of pixels between two patches in both x and y dimensions. Setting a higher value of step will reduce computational cost but it will also degrade the reconstruction.

Example

     Xhat = BPFA(Y, mask, 15)
     Xhat = BPFA(Y, mask, 15, 'iter', 50)
     Xhat = BPFA(Y, mask, 15, 'iter', 50, 'step', 3)

Python code usage

Function file: BPFA_for_Python/BPFA.py

First note

To have access to the BPFA code and to be able to import it, add its path to the PYTHONPATH.

As an example, in Linux, you can add the following lines to your .bashrc:

export PYTHONPATH="$PYTHONPATH:/home/path_to_module/BPFA_for_Python/"

Arguments

Name	Type	Description
Y	2D or 3D Numpy array	The input data.
mask	2D Numpy array	The acquisition mask.

Keyword Arguments

Name	Type	Description
PatchSize	int	The patch size. Default is 5.
Omega	int	The Omega parameter. Default is 1.
K	int	The dictionary dimension. Default is 128.
iter	int	The number of iterations. Default is 100.
step	int	The distance between two consecutive patches. Default is 1 for full overlap.
verbose	bool	The verbose parameter. Default is True.

Returns

Name	Type	Description
Xhat	2D or 3D Numpy array	The reconstructed image.
InfoOut	dictionary	Dictionary containing the following informations: 'dt' which is the elapsed time.

Example

     import BPFA
     Xhat, infoOut = BPFA.BPFA(Y, mask, PathchSize=15)
     Xhat, infoOut = BPFA.BPFA(Y, mask, PathchSize=15, iter=50)
     Xhat, infoOut = BPFA.BPFA(Y, mask, PathchSize=15, iter=50, step=3)

MATLAB and Python versions

The repository code has been developped on MATLAB 2017 and Python 3.5.

Author and License:

This code was improved by Etienne Monier (https://github.com/etienne-monier) based on a code developed by Zhengming Xing, Duke University (https://zmxing.github.io/).

The previous code did not have any license. Then, this repository is shared under MIT License.