
# Building a dictionary¶

This tutorial explains how to generate a dictionary required by PyHST2 when using the Dictionary-based reconstruction method. A typical use case is when you have a high quality reconstructed slice of a sample, and you want to perform low-dose scans on similar samples afterwards. The dictionary will capture the essential features of the high-quality slice, and will be used to reconstruct other volumes.

## Installing ksvd¶

You first need to install ksvd, a Python script which generates a patches file from a slice/volume. In this tutorial, the directory /home/myself/software/ksvd is used as an example.

mkdir software/ksvd
cd software/ksvd
export KSVD_DIR=$(pwd) git clone http://gitlab.esrf.fr/mirone/ksvd cd ksvd python setup.py install --prefix=$KSVD_DIR


Then, set the environment variables. This has to be done each time you want to run ksvd.

export PATH=$PATH:$KSVD_DIR/bin
export PYTHONPATH=$PYTHONPATH:$KSVD_DIR/lib/python2.7/site-packages/


## Running ksvd¶

Do not forget to set the environment variables :

export PATH=$PATH:/home/myself/software/ksvd/bin export PYTHONPATH=$PYTHONPATH:/home/myself/software/ksvd/lib/python2.7/site-packages/

The syntax is ksvd input_file.yaml. You need :
• An input file

• A slice or volume from which the patches will be built

## Input file format for a slice¶

ksvd uses an input file format in the .yaml format. Suppose you have a (good quality) slice (in EDF or vol format) myslice.edf from which you want to build the dictionary patches. A basic input file will be :

patchwidth: 8
NPURSUIT:    4
n_comps:   200
nitersksvd:  5
files:
- !ImageFile
START_X: 0
END_X: 100000
START_Y: 0
END_Y: 100000
VAL_MIN: -1.0e+30   # no constraints
VAL_MAX: +1.0e+30   # no constraints
FILE_TYPE: edf
FILE:  myslice.edf
DERIVATIVES: 0
N_PATCHES: 40000


This example produces a dictionary of 200 $$8 \times 8$$ patches.

You can build the patches from a region of interest with the parameters START_X, END_X, START_Y, END_Y. If the entire slice is used, you can set these parameters to a huge value.

You can also discard values outsides of the range [VAL_MIN, VAL_MAX].

## Input file format for a volume¶

The input file for a vomume is slightly different. A example of file building patches on the volume myvol.vol is the following :

patchwidth: 8
NPURSUIT:    4
n_comps:   200
nitersksvd:  5

files:
- !VolumeFile
START_X: 600
END_X: 1428
START_Y: 600
END_Y: 1333
START_Z: 31
END_Z: 51
VDIMX: 2048
VDIMY: 2048
VDIMZ: 81
VAL_MIN: 0.1
VAL_MAX: 0.29
FILE_TYPE: vol
FILE:  myvolume.vol
DERIVATIVES: 0
N_PATCHES: 40000


In this example, 200 $$8 \times 8$$ patches are generated from the region of interest $$[600, 1428] \times [600, 1333]$$. The slices 31 to 51 are used in the volume. The parameters VDIMX, VDIMY, VDIMZ specify the total dimensions of the volume.

## Input file format for vectorial patches¶

The Dictionary-based reconstruction can exploit the 3D correlation within the volume (instead of the 2D correlation in a slice). To do so, the patches need to be three-dimensional (hence they have to be learned on a volume).

The following example generates 300 patches of size $$8 \times 8 \times 3$$. This means that in the reconstruction, the slice will be reconstructed by groups of 3. Here, the provided volume has 7 slices, which is enough for this example.

patchwidth: 8
NPURSUIT:    4
n_comps:   300
nitersksvd:  5
patch_eight: 3

files:
- !VolumeFile
START_X: 0
END_X: 1163
START_Y: 879
END_Y: 1333
START_Z: 0
END_Z: 6
VDIMX: 2048
VDIMY: 2048
VDIMZ: 7
VAL_MIN: -1000.0
VAL_MAX: 1000.0
FILE_TYPE: vol
FILE:  myvol.vol
DERIVATIVES: 0
N_PATCHES: 40000


## Input file for differential phase contrast¶

PyHST2 has a reconstruction mode for differential phase contrast (eg. analyzer-based phase contrast). In this case, each line of the sinogram is multiplied by $$\cos(\theta)$$ and $$\sin(\theta)$$, which gives two sinograms, and eventually two slices. The slice corresponding to the $$\cos (\theta)$$ component will be reconstructed with a set of patches, while the slice corresponding to the component $$\sin (\theta)$$ will be resconstructed with another set of patches.

Thus, the dictionary contains patches with two components. To build this kind of patches, the option to set is

DERIVATIVES: 1


## Visualizing the patches¶

You can also visualize the patches. In the directory where your patches file is located, just type

ksvd


(without arguments), and then provide the name of the patches file when prompted.

## A common error¶

If you do not have the version 5 of PyMca, you will get the following error :

Traceback (most recent call last):
File "/home/myself/software/ksvd/bin/ksvd", line 3, in <module>
import ksvd.ksvd
File "/home/myself/software/ksvd/lib/python2.7/site-packages/ksvd/ksvd.py", line 7, in <module>
from PyMca5.PyMcaIO import EdfFile
ImportError: No module named PyMca5.PyMcaIO


You need to modify the lines 7-8 of the file /home/myself/software/ksvd/lib/python2.7/site-packages/ksvd/ksvd.py. The lines

from PyMca5.PyMcaIO import EdfFile
#from PyMca import EdfFile


should be

#from PyMca5.PyMcaIO import EdfFile
from PyMca import EdfFile