In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010
detection. To accurately describe the straight edge segments, a
principal component analysis technique is adapted. To establish the
line to line correspondences between the stereo images, a new pair-
w’ise stereo matching approach is developed. The approach involves
new constraints and a final probabilistic elimination to reduce the
number of false matches.
2.1 Pre-Processing
2.1.1 Multi-level Non-Linear Color Diffusion Filter: The aim
of the non-linear diffusion filtering is to eliminate the noise inherent
in the images without blurring the step discontinuities. A good
review related to diffusion filters can be found in (Weickert, 1997).
Since we deal with multispectral images, it is impoitant to perform
the filtering procedure considering those discontinuities in different
bands. In order to achieve this, we adapted a gradient computed
through tensor mathematics (see section 2.2.1) to improve the
performance of the original non-linear filter. In addition, we utilized
a three level processing chain (decreasing the sigma parameter
w'hile increasing the lambda parameter) to diminish the noise
around the step discontinuities.
2.1.2 Color Boosting: The goal of color boosting is to improve
the apparent color difference between adjacent objects in a scene.
For the aerial images (especially for analog cameras), the contrasts
in the RGB values caused by the color variations are generally not
high enough to exploit this distinction. Therefore, the idea is to
amplify the color variations between the objects (for example, a
building roof and its background) before the edge detection to find
the edges that cannot be detected due to low' color variation. We
utilized the boosting technique developed by Weijer et. al. (2006a).
First, the RGB color space is transformed to the decorrelated
Opponent Color Space (ol, o2, and o3). Next, to improve the color
contrast in the images, color directions of the opponent space (ol
and o2) are selected and multiplied with a factor of k (k>l). Finally,
the modified opponent color space is back-transformed to RGB
color space.
2.2 Line Extraction
2.2.1 Color Edge Detection: In this study, to maximize the
performance of the edge detection, we utilize the algorithm
developed in Weijer et. al. (2006b). They proposed a color Canny
edge detection algorithm to locate the edges accurately in
multispectral images. The algorithm mainly consists of the
calculation of the spatial derivatives of the different image
channels, and the computed derivatives are combined using tensor
mathematics. In this way, differential structures of the bands in
multispectral images are mutually supported, so that edge detection
of better completeness is accomplished compared to the single band
detectors. Two minor adaptations enhance the results of the
algorithm: (i) the output of the final gradient map is scaled between
zero-and-one before further processing, which significantly reduces
the remaining noisy edges, and (ii) a two level hysteresis
thresholding is designed to have a better control on the final edge
contours (Fig. lc, Id).
2.2.2 The Extraction of the Straight Line Segments: We offer
a two stage solution to the straight line extraction problem, (i) the
extraction of straight edge segments, and (ii) robustly fitting line
segments to the extracted straight edge segments.
We use the principal component analysis technique developed to
extract the straight edge segments. The details of the method can be
found in Lee et. al. (2006). Although the method has proven to be
more efficient in several w'ays than Flough Transform (Lee et. al.,
2006), w'e experienced several problems during the extraction of the
straight edge segments. First, the input binary edge images are
assumed to be segments that are only a single pixel wdde. However,
this is not the case for the output of the binary images generated by
(e) (f)
Fig. 1 (a, b) Test images from Vaihingen, Germany, (c, d) the
results of the color Canny algorithm, (e, f) straight edge segments.
the color canny edge detection. Although non-maximum
suppression is applied after the detection stage, this does not always
guarantee one pixel wide edges for color images, since separate
spatial derivatives of the image bands are combined during edge
detection. In this study, we utilized the image skeleton technique to
remove the redundant boundary pixels of the binary edges. The
technique ensures that the binary objects shrink to a minimally
connected structure without breaking apart. A different critical
shortcoming we observed is that, if two same label (for example
two column-directional) binary edge segments are connected with a
junction of a narrow angle, the algorithm is no more capable to
determine the correct straightness value. Unfortunately, this type of
line to line combinations is not rare in aerial images. To solve the
problem, w'e identified all potential line to line (or multi-line)
endings and crossings within each segment. Thereafter, the
problematic crossings of the edge segments are removed.
We refer to a line segment as a single straight object that is
composed of only two endpoints (xi, yi; x^, ya). To accurately
describe the line segments, in this study, the well-knowm Ransac
algorithm is utilized (Fischer and Bolles, 1981; Zuliani et. al..
2005). In some cases, a single straight edge segment may be
represented by more than a single straight line. For those cases, a
recursive strategy is applied to describe each line segment from the
straight edge segments. Fig. 3a and 3b illustrates the line segments
extracted for the images given in Fig. la and lb.
2.3 Pair-wise Stereo Line Matching
Once the straight lines are extracted, a matching strategy is required
to find the line correspondences between the reference and search
images. We propose a new pair-wise stereo line matching strategy
that consists of two fundamental stages: (i) selection of line pairs on
the reference image, (ii) identifying the candidate pair models on
the search image.
