1. Introduction to Image Processing
EN
Image processing is known as a separate scientific discipline
since the early sixties and began its development on the basis
of the mcthods being used in character recognition. Figure 1
shows the block diagram for an automatic digital image processing
system and the different units will shortly be discussed within
this introduction.
The input to an image processing system is a natural scene or
a photograph recorded on film or paper. This input information
has to be digitized in the scanner and the result is usually
stored on magnetic tape or is directly transferred to tho me-
mory of a digital computer. Photographs are typically resolved
in different colour channels with up to 2000 by 2000 picture
elements (pixels) in 8 bit grey levols covering an area of
about 60 by 60 mm. According to the accuracy and resolution
being required rotating drums, scanning tables which can be mo-
ved in two directions, TV systems, image dissector tubes or lignt
deflection systems are used.
In many applications a scene (e.g. the surface of the earth) is
directly scanned without recording on film using multispectral
Scanners. For remote sensing tasks sensors with different spec-
tral sensitivity and active or passive sensing principles are
applied covering a broad field of measuring problems, such as
sensing the earth from an aircraft using synthetic aperture
radar (SLAR).
For a quick interpretation of the results or for monitoring the
different processing steps it is often necessary to generate a
twodlimensional representation of an array of grey values stored
in the computer. For this purpose displays with storage or non-
storage tubes or recording systems for the exposure of photogra-
phic materials are used. The problems related with the Scanning
and generating of images have been solved for most of the inter-
esting applications, a lot of systems are commercially available.
The next step in image processing is called preprocessing. By the-
se preprocessing algorithms the image is usually transformed to
another twodimensional representation for the purpose of geometric
and radiometric manipulations (figure 2). The geometric propro-
cessing may include the generation of graphic overlays of the image,
the rectification for eliminating distortions produced by the scan-
ner or the geometric match of several images for the purpose of
change detection. The radiometric preprocessing is done to decrea-
se the distortions produced by scanning and digitizing the origi-
nal data. An other broad field of radiometric preprocessing is
the enhancement of images to decrease the amount of noise or to
produce a better representation of the interesting objects. Accor-
ding to figure 3 the radiometric preprocessing can be done in the
image domain or in the frequency domain (e.g. Fourier domain).
Some of the methods are based on single pixels, that means the
greylevel of a pixel in the preprocessed image depends only on
the value of the corresponding pixel in the original image. But
most of the algorithms use local operations considering in addi-
tion the local neighbourhood of the pixel. The operator used may
be linear or nonlinear and homogeneous (the same operator for
the total image) or non-homogeneous (the operator deponds on the
information or position of the pixel within the image). In the
frequency domain it is only possible to use methods based on line-
ar and homogeneous algorithms.
One example for using preprocessing methods is change detection
where changes in two similar images of the same region have to
be recorded. After a geometric and radiometric match of the two
images the grey levels of corresponding pixels are subtracted.
Using a threshold to eliminate the amount of noise the changes
in the images may be displayed for further interpretation.
After the preprocessing the image has to be analysed in order to
find the appropriate features of the interesting objects. Figure
4 shows a classification scheme of possible feature extraction
methods. The pixel oriented methods generate one feature vector
