Image Processing Reference
In-Depth Information
9.3 Results
The soft morphological filter used in this work was designed with a genetic algo-
rithm.
9
The nonlinearities in the filter make it difficult to produce a deterministical-
ly designed optimum solution. Instead, an iterative search approach is used that
tests a number of different solutions. It combines and modifies these solutions until
no further improvement is possible.
The training and application of the soft morphological filter was performed in
three steps:
1.
A training set was created that mimicked the effect of the disturbance in the real
data.
2.
Training was performed using a genetic algorithm (GA) and the improvement
using training data was confirmed.
3.
The resulting filter was applied to the real data and the improvement was ob-
served.
9.3.1 Creation of a training set
The first step in removing the noise from any image is to understand the nature of
the disturbance. In this case, the distortion was caused by cosmic rays hitting the
CCD in the SOHO telescope and causing the cells to overload, producing an image
that suffers from extreme “white out.” In many cases, the obstruction was severe
enough to render the data worthless in its current form. It was vital to ensure that the
noise model used in training was appropriate to that affecting the real images. A
poorly chosen noise model will lead to a poorly trained filter.
In order to create a filter to remove this noise, a method of filter training was re-
quired. A genetic algorithm was used to determine the optimum filter applied to a
training set of representative ideal and noisy image pairs. The noisy image was fil-
tered and the output compared with the ideal image. The GA was used to adjust the
filter parameters iteratively in order to make the output as close as possible to the
clean image.
The representative training set was created using ideal images, and adding
noise from the real images to create a set of noisy images. For this example, clean
images from the SOHO telescope were taken from sohowww.nascom.nasa.gov.
These images were cropped from 1024
×
1024 pixels to 150
×
150 pixels to reduce
training time.
The iterative nature of the filter design can result in large processing times. It is
important that the balance between training set size and the search space of the fil-
ter, as discussed in Chapter 4, is maintained. It was found from experience that us-
ing ten cropped images produces well-trained filters. Once the images were resized,
white patches of speckle noise, similar to those seen in Fig. 9.2, were added to cre-
ate the corresponding noisy images. This was done manually through a cut and
paste operation. Together the two sets of images formed the training data.
