• No results found

1 . In the remainder of this thesis methods relating to the improvement of the "educational environment" are developed and described. The methods of improving the pupil have been briefly sumarised above and should, in general, be considered as an adjunct rather than an alternative to the work which is described in the following chapters.

2. The experiments described in this chapter represented a preliminary study of a neural network based universal window filter. They show that the NNWF

can learn filtering operations from representative input-target image pairs and, at least in the situations described, generalise the required operation. These results suggested further study of this technique was warranted. Very few problems with local minima were encountered. In general

if

a network did not converge to an acceptable minima restarting with a new set of random weights did not help.

3 . The small size of the networks which were required to perform these transformations suggest that a hardware implementation of the

NNWF

could be practical. A relatively small number of inputs and hidden nodes are used

in these networks. The most complex network used routinely in the work described in this thesis had only a 9x9 input window and two hidden layers containing five and three neurons respectively. This had the consequence that simulations could be performed in a reasonable time on the computer systems available.

4 . I f the filter network described i s to be considered as a universal window filter the creation of targets needs further consideration. How can target images be provided in the more general case where an existing algorithmic technique is not being emulated? A number of techniques can be considered;

creation a domain

This could be done entirely by hand or by using a variety of image processing and computer graphics techniques to construct the target. The capability of the

NNWF

to learn subjective distinctions should be an advantage in this situation as it takes advantage of the non-algorithmic nature of neural networks. A fmal system can be envisioned in which a

Chapter 3 A Universal Neural Network Based Window Filter 3.38

sample image or part image would have the features of interest marked o n it by a domain expert. These image pairs would be· used to train the

NNWF filter. The remammg images would then be classified automatically.

An image processing expert could be used to select and apply a sequence of operations to a sample image to provide a suitable target image for training. This may either be constrained to only include non-subj ective steps or may include subjective evaluations (for example choosing a threshold level which best suits this particular image.

Using a slower or more complex image capture or rmage creation procedure to produce a required target. For example

if

the NNWF is to remove noise due to the short integration time which is required in a particular situation then it may be possible to provide extra lighting or multiple exposures for the construction of the target. Naturally

if

the image can be improved easily and routinely in this way then this would be preferable to later filtering with the NNWF. It is only in the case where the extra effort or time can be j ustified only during the training that this technique is applicable.

creation the normal flow.

In some situations it is possible to reverse the normal flow of operations and produce the input image from the target image. For example, starting with a sharp image which will be used as a target a blurred, or noisy input image can be created. The NNWF could then be trained using this pair to remove blur or noise. The experiments in section 3.3 above provide an example of this method of training the NNWF.

5 . Even though the emulations of the window filters were successful they were not perfect and in the case of the Marr-Hildreth considerable improvement could be sought. The two main approaches to improving the performance are; changing the network and changing the training. The first of these has

Chapter 3 A Universal Neural Network Based Window Filter 3.39

already been discussed in section 3.4 the second is developed in chapter five and later chapters of this thesis.

6. Other possible extensions to the basic structure of the NNWF include;

• input windows which are non-square • multi-resolution input windows

• augmenting the inputs with other types of information such as image statistics or window position.

• Using parts of the output image already scanned to augment the

network input.

While this work is not drawn on in the rest of the thesis some preliminary experiments on the last three extensions mentioned above has been done. This is briefly described here;

• The NNWF was modified to incorporate a multi-resolution window in

which a central 3x3 pixel window was surrounded with a ring of eight regions each region being 3x3 pixels, This was in turn surrounded by a ring of regions each 9x9 pixels in size. The intention of this arrangement was to provide fine detail information in the centre and then gradually more general or averaged information as the distance from the centre of the window increases. A NNWF of this form was implemented and was able to learn filtering operations. This form of the NNWF has not been applied to any significant tasks as yet.

• In order to enable the NNWF to perform operations in which the position

within the image should influence the output the NNWF network inputs were augmented with the x and y coordinates of the centre of the window. This was used to train the NNWF to remove image distortion from a set of similarly distorted images.

• To allow the NNWF to learn to favour the production of continuous lines

in the output images the NNWF was modified to augment the network inputs with the pixel values in a corresponding window in the output image. The window used did not include values from below or to the right of the central pixel. Only initial experiments have been performed with this NNWF implementation.

Chapter 3 A Universal Neural Network Based Window Filter 3.40

The investigation of non square windows and further exploration of the variations discussed above has been left as possible future work. It was decided at this point in these investigations to concentrate research on improving the educational environment. Other areas worthy of further investigation include; finding methods of improving the network( or pupil), extending the basic structure of the NNWF and pro�ucing parallel or hardware implementations of the network filter. These all involve interesting possibilities but were not pursued as part of this thesis.

Chapter 3 A Universal Neural Network Based Window Filter 3.41