347
346
III.l] Input images
The concepts and quality indicators can be differentiated into
deterministic and statistical and further into those referring to
image quality (i.e., pictorial) and to spatial location (i.e., geo-
metric). :
Image quality: deterministic concepts (and tools) are Fourier Trans-
form (or power spectrum), dynamic range, resolution (spatial, inten-
sity, spectral), and modulation transfer-function (MTF-imaging).
Common statistical concepts are signal-autocovariance (and varian-
ce), intensity hystogram, spread function, noise-autocovariance (and
variance), signal-to-noise ratio (SNR), and noise-power spectrum.
Geometry: Deterministic concepts include the different models of
regular distortions, such as for film, camera, and atmosphere.
Statistical concepts (and indicators) comprise weight functions,
covariance (or correlation) functions, variances, and various sta-
tistics (error distributions), all referring to geometric location.
III.2 Digital raw image data
The concepts and quality indicators can be classified as those for
input images.
Image quality: Deterministic indicators are intensity increment (or
resolution), spatial increment (and pixel size), Fourier Transform
(FFT and/or power spectrum of digital image), and Transfer Function
(of rastered image) and corresponding limiting frequency Zl [6].
Statistical measures are autocovariance (and variance) of digital
image, level of noise, SNR, and different statistics (distributions)
Geometry: Deterministic indicators are model deformations caused by
relative orientation, and errors in positioning sensors at sampling..
Statistical quality measures are standard deviations caused by spat-
ial incrementation (1.e., image raster : g - A:3.46, where A is
increment size; [2]y and locational-error statistics.
III.3 Upgraded image data
The concepts and quality indicators should reflect the specific
choice of the pre-processing operations, which can vary from one
case to another.
Image quality:
The degrading effect of data resampling can be assessed by a compo-
sed Transfer Function (annex 1), whereas that of linear filtering
can be handled by a single Transfer Function r2. Thresholding impo-
ses limits on either intensity or frequency for selective data
exclusion (annex 2). Corresponding losses are quantifiable, e.g., in
terms of standard errors. Image transformations (intensity) are
accompanied with losses too; the corresponding assessment, however,
is beyond the scope of this paper. Synthesising image data does not
introduce significant losses if done with sufficient care.
Failures and errors in image analysis and feature extraction are
hardly tangible.