The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
included attempt at automatizing the first stage of the analysis,
or checking whether it is possible to state that the selected
image fragment is "informative" enough and can be processed
further.
4.1 Sub-image representations
Proposals for creating a sub-image representation were shown
in publication (Mikrut et al, 2008). Three methods have been
implemented and tested: an algorithm based on a two-
dimensional histogram of gradients, log-polar and log-Hough
transforms, as well as signatures generated by the pulse-coupled
network ICM.
The method described "histogram of angles" consists in making
a two-dimensional histogram of a gradient image (Mikrut et al,
2008). A single field of that histogram corresponds to the
number of pixels, for which the values of the gradient and of the
direction fall within the range that has been set. A two-
dimensional histogram of the gradient is prepared for all those
image pixels, which are edge pixels. Based on the so-prepared
histogram, a vector of features is determined in the form of a
profile of the maximum and average values of the gradient
module, which are generated through projecting maximum and
average values, respectively, upon the gradient module axis,
and a profile of the maximum and average direction values.
Details of the method can be found in the referenced work
(Czechowicz et al., 2007). The so-created representations
provided input for the self-organizing Kohonen neural network.
That network was used for the selection of images, and the
selected neurons of that network (or rather their output signals)
formed the second type of representation, which was based on
the 2D histogram (see Fig. 3 on the left).
The log-polar and log-Hough transforms briefly described in
Chapter 3 made use of a non-processed image. The feature
vector was created through projecting the maximum values
from the respective lines and columns in the log-Hough space
upon the coordinate system axes. In that way profiles were
generated, which determined the position and length of straight
sections and gentle arcs of the edges. Information on the
sequence of image transformations and on the selection and
aggregation of feature vectors were published in the paper
(Piekarski et al., 2007).
The last type of representation were sub-image signatures,
generated by the ICM pulse-coupled network. The experiments
involved signatures made up of 25, 50, and 100 elements
(Mikrut, 2007).
4.2 Image selection
The task of selection was based on the use of neural networks
for classifying sub-images as "favourable" and "unfavourable"
ones from the point of view of their subsequent matching. The
main database was prepared on the basis of aerial images of the
area of the city of Krakow in grey scale (G component from the
RGB colour image), with the resolution of 600 dpi, from which
900 sub-images were selected, each of the size of 240x160
pixels. Five experts classified each of those sub-images
independently as belonging to one of three groups: favourable,
unfavourable, and intermediate areas in respect of seeking out
elements for image matching. The median value of
classification determined that a given sub-image belonged to a
specific class. The samples were divided into two sets: the
teaching and the testing one.
Two types of neural networks were used for the selection:
Kohonen's SOM and backpropagation networks (see Fig. 3).
Both the analysis of the database and preliminary experiments
showed that results obtained for the three classes were not
satisfactory. That is why in further tests neural networks were
taught to make a division into two classes: the "favourable" and
"unfavourable" sub-images.
Finally, input for the Kohonen network was provided by 19-
element representations, made of profile of average gradient
direction value, with the 20° aggregation. The best results were
obtained using a network of a size of 7x4 neurons.
In the other three cases (cf. Fig. 3), backpropagation networks
were applied. The experiment methodology was similar in each
case. At first, the pairs of "representation - network structure"
were roughly determined, and then the specific network, that
had yielded the best recognition results in the initial phase, was
taught by means of the selected representation. Using the
rejection technique increased the efficiency of recognition. The
most effective network structures and percentage classification
results are specified in Chapter 5.
4.3 Sub-image representation matching
Two sub-sets were selected from the database described in
Section 4.2: 36 images representing the G component, and 44
images in greyness degrees (the R, G, and B components have
been integrated). Those areas were manually identified on
another aerial image and, after proper framing, they were
recorded to perform matching tests.
It should be stressed that the research works under discussion
involved testing the possibilities of sub-image matching, which
means that sub-image representation vectors from the first
(chronologically) image were computed, then representations of
all sub-images in the second aerial image fragment were
computed, and those vectors were compared. The actual
comparing of vectors was effected with the use of the classic
correlation.
As the bottom right part of Figure 3 shows, there were-three
types of representations subjected to matching: representation
obtained with the use of Kohonen's SOM network, signatures
generated by the ICM pulse-coupled network, and - for
comparison - vectors of the gradient histogram that were
obtained without the use of neural networks.
To test matching, a 19-element representation of profile of
average gradient direction values with the aggregation of 20°
was selected. Responses generated on the best classifying SOM
network of the structure of 19-28[7x4] constituted the neuron
representation. Details can be found in (Czechowicz et al.,
2007a), and the most significant results were tabled in Chapter
5.
The other representation, which was used for matching,
included 100-element signatures generated by the ICM network.
Results of matching in relation to signatures generated on
images that have been two and four times decreased in size
were investigated. Signatures for the whole image, as well as
for the image divided into 4- and 6-element fragments by means
of a uniform grid were computed. The description of those
experiments will be published in (Mikrut et al, 2008) while the
best results are presented in Chapter 5.
334