331
NEURAL NETWORKS IN THE AUTOMATION OF PHOTOGRAMMETRIC
PROCESSES
S. Mikrut 3 '*, Z. Mikrut b
a AGH, Dept, of Geoinformation, Photogrammetry and Environmental Remote Sensing, University of Science and
Technology Cracow (AGH), 30059 Krakow, Poland - smikrut@agh.edu.pl
b AGH, Dept, of Automatics, University of Science and Technology Cracow (AGH), 30059 Krakow,
Poland - zibi@agh.edu.pl
Commission IV, WG IV/3
KEY WORDS: Photogrammetry, Automation, Correlation, Extraction, Recognition, Neural
ABSTRACT:
The concept of research was based on the selection of several representations, which later were correlated by means of classic, and
neural methods. In the course of research, classic methods of image matching were tested and compared with neural methods that
originated in the course of research. Additionally, experiments consisting in manual measurements, performed by independent
observers, were conducted. The essence of methodology that was based on neural networks consisted in the preparation of suitable
representations of image fragments and using them for the classification of various types of neural networks. One of the assumed
methods was based on the distribution of image gradient module value and of its direction. The usability of that representation for
the selection of sub-images was tested by means of SOM Kohonen neural network. Another method consisted in the utilization of
the log-polar and log-Hough transforms, which are considered to be simplified models of preliminary image processing, performed
by visual systems of people and animals. The usability of that representation was tested by means of the backpropagation type of
neural network. As regards the generation of the third representation, the ICM (Intersecting Cortical Model) network was applied,
which is one of the versions of the PCNN (Pulse Coupled Neural Network). Using that network, the so-called image signatures, or
vectors composed of tens of elements which describe the image structure, were generated.
1. INTRODUCTION
The use of neural networks to solve photogrammetric problems
is a relatively new direction of research, although those
networks have been used in digital image analysis for long. The
problem consists not only in the lack of popularization of that
technique among photogrammetry specialists. An additional
difficulty results from the specificity of photogrammetric
images, which are marked by a high diversity and
fragmentation of information. The smaller is the scale of images,
the greater is that fragmentation. This gives origin to troubles
with the creation of a set of data for teaching and testing, as that
set should, first of all, be representative. Another difficulty is
the lack of neural tools, with which one could perform one of
the most significant stages of the analysis, that is matching
(correlating). Therefore, there remain issues related to the
classification of areas and detection of features, as shown by the
application examples drawn from literature and presented in
Chapter 2.
The authors of this paper have proposed a conception of using
neural networks for the preliminary selection of sub-images and,
indirectly, for matching them. The paper is a recapitulation of
several years' works. The research conception was based on a
selection of several representations, which were later used as
inputs for the classifying neural networks. Broadly understood
neural networks were also used to create representations, which
were then correlated. The classic methods of image matching
were tested and compared with neural methods, which
originated in the course of research. Additionally, experiments
consisting in manual measurements performed by independent
observers were conducted.
The arrangement of the paper is the following. Chapter 2
presents examples of the application of neural networks in
photogrammetry and in related issues taken from the literature
on the subject. Chapter 3 recollects principles, associated with
the use of neural networks technique, as well as discusses to a
greater extent a less known ICM (Intersecting Cortical Model)
pulse-coupled network. The main part of the paper is included
in Chapter 4, which presents the authors' ideas related to the
feasibility of application of neural networks on different stages
of photogrammetric image analysis. It also discusses briefly
those concepts, which have been implemented. Chapter 5
includes the specification of results and conclusions.
2. EXAMPLES OF THE APPLICATION OF NEURAL
NETWORKS IN PHOTOGRAMMETRY AND REMOTE
SENSING
The attempts to use neural networks to solve various problems
in the fields of photogrammetry and remote sensing have been
made, practically speaking, from the very beginning of their
development. The possibilities of using an artificial intelligence
in photogrammetry were presented as early as in 1988 (in a
general outline) by T. Sarjakoski (Sarjakoski, 1988).
One of the first applications concerned the use of networks for
feature extraction (Zheng, 1992). At the same time, Hu (during
Corresponding author.