| The opening unit:
Here, a new image can be opened. The tool supports the following file
formats: jpeg, tiff, gif and pong, in truecolor (RGB),
indexed color, and gray levels. Images of type uint8
are converted to doubles, and normalized between 0 and 1 by dividing all
pixels by 255. Index color images are converted to truecolor images. The
'view image' button opens a new figure window and displays the image.
The Gabor filtering unit:
Log radial Gabor filtering is applied to have the
texture information. Gabor filteres are 2D Gaussian bandpass filters
defined by a center frequency and a bandwidth. For 2D images, both the
center frequency and the bandwidth must be defined in two directions.
Radial Gabor filtering uses polar coordinates to align the Gaussian. See the contour plots above. The center frequencies and
the principal axes of
the filters are aligned along the radial direction, and along the
angular direction. Log radial Gabor uses a logarithmic scale for the
radial frequency. That is, on a log frequency axis, i.e. log(r), the
transfer function is a Gaussian:
The radial frequencies are normalized
such that a frequency of 1 corresponds to half the sampling frequency.
In the gui shown above, the highest peak frequency (is the centre frequency
of the filter that passes the highest frequencies) is set to 0.8. The
ratio of neighboring peak frequency is 0.5. Thus the sequence of center
frequencies is 0.8, 0.4, 0.2, 0.1, 0.05, ... However, the number of
filters in radial direction is 3. Therefore, these filters have centre
frequencies of 0.8, 0.4 and 0.2, respectively. The relative bandwidth,
s, is
0.65, expressed on a logarithmic scale. On a linear scale of r, this
actually implements a deviation that is proportional to the center
frequency.
The figure above specifies 6 filters in
the angular direction implying that the difference of neighboring angles
is 180/6=30 degrees. The relative bandwidth in angular direction is 0.667
meaning that the standard deviation of the filter transfer is 30*0.667=20
degrees.
For each filter the inphase and
the quadruture component-images are calculated. The magnitude of
these components is the final result. For each filter, the magnitude is
stored as an image (with the same number of pixels as the input image).
For color images, this procedure is applied to each color-band image, i.e.
to R, G, and B. The multiple images that result from this procedure are
stacked into one vector-valued image. In the example above, there are
3*6=18 filters. Thus, the output image (referred to as 'Gabor images') has
18 feature elements per pixel. In case of a color image this would be 3*18=54
elements.
With the button 'view Gabor image', the
output of one of the filters (i.e. a Gabor image) is displayed in a new
figure window. The amplitude of the images are scaled so as to match the
range of the graphical display adaptor.
Note that if the input image shows a
modulated harmonic function, then the Gabor filter that is tuned to that
harmonic function produces the envelope of the input image, i.e. the
modulation signal.
Selection of pixels:
There are three modes of pixel selection:
- The full image (i.e. all pixels are
selected).
- A random selection
- Selection of the pixels that are on a
regular orthogonal grid. The variables 'deltaH' and 'deltaV' denote the
pixel interval of the grid in horizontal and vertical direction.
The 'random selection' and 'selection on
a grid' are added in order to provide a facility to tam the number of
feature vectors. Without these modes, the training data can become
unmanageably large.
A second facility for the selection is to
use a ROI (region of interest). With the button 'define ROI' the user can
define a polygonal ROI. If the corresponding radio button is on, then only
pixels inside this ROI will be selected. This facility offers the
possibility to create a labeled data set. The default ROI is the whole
image area.
The output section:
The output section consists of three parts:
- In the first part, the ROI can be
saved in an image format (jpeg, tiff, gif, or pong).
- In the second part, one or all of the Gabor images can be written to file(s)
in an image file format (jpeg, tiff, gif, or pong).
- The third part can be
used to write the created feature vectors of the selected pixels to a file
using Matlab's 'save' command. The output file has the *.mat file
format. Therefore, the data can be retrieved by using Matlab's 'load'
command. The vectors are stored as row vectors in a matrix (with the same
name as the filename of the mat-file). Thus, if 1000 pixels are
selected, and the feature vector is formed by 18 Gabor images, then the
size of the matrix is 1000*18. If the vectors are saved to the file 'foo.mat',
then loading this file will create a matrix in the workspace, named 'foo'.
Version compatibility:
ut_getgabor is developed with Matlab
6.5sp1
Author:
Copyright (c): F. van der Heijden, all
rights reserved, November 2004. |
|
