The general version of resampling is mixed, however. A new image
raster is created, having a different density (cell size) and orien-
tation than the original raster. A specific case is the epipolar
lay-out of pixels.
Resampling comprises two basic stages, i.e., calculation of the new
pixel locations (geometric lay-out) and interpolation of the corres-
ponding intensity levels from the neighbouring old pixels. Resam-:
pling is accompanied with some loss of pictorial data; it should
therefore not be applied cyclically in an iterative process.
5. Data segmentation and structuring
Image data segmentation and their structuring are interelated, and
they strongly reflect the strategy of image matching. Segmentation
and structuring can refer to the spectral domains (spatial or elec-
tromagnetic), to spatial entities (i.e., lay-out of patches, seg-
ments and pixels), or to a combination of both.
Lay-out of the spatial data entities (patches, segments) can be
fixed (and homogeneous) or variable, e.g. adapted during pre-proces-
sing or matching. Data entities can be arranged in a single level
or in several hierarchical levels (2).
A homogeneous lay-out of data entities can be single or multiple. In
the latter case segments can be shifted in x and/or y, (e.g., for
half interval),changed in size (which changes overlap of segments),
etc.
Further variation is attained by interchanging the roles of target
and search segments (i.e., LH versus RH image). Moreover, segment
size (of target) can be fixed or variable. In the latter case, it
can be adapted during pre-processing or during matching.
6. Image transformations
Pictorial (intensity) transformations and matching algorithms are
strongly interelated; a transformation may be regarded as an inte-
gral part of matching. Examples are Fourier, Hadamard, and exponen-
tial transforms, which can serve for image analysis and segmenta-
tion. Transformations are accompanied with some loss of image data.
7. Filtering and thresholding
Linear filters are commonly applied in order to compensate for de-
gradation at imaging (by inverse filters) and to suppress noise (by
smoothing filters). Filters and thresholds can be aplied in selec-
tive pre-processing, i.e., to enhance certain image features, for
segmenting and/or compressing image data, for interpolation, etc.
Different linear filters can be merged (by convolution) into a com-
posed filter, which increases efficiency of processing.
8. Synthesising image data
The reliability of iamge matching can be increased by creating syn-
thetic image data from the original (or pre-processed) data. Exam-
ples are envelope data (3), compressed second difference data (2),
etc.
336