New Saliency Point Detection and Evaluation Methods for Finding Structural Differences in Remote Sensing Images of Long Time-Span Samples - Advanced Concepts for Intelligent Vision Systems - page 276

The keypoint candidate detection method was the same as written before.
Results are in Figure 4(b). Keypoint candidates cover all buildings, and only a
few points are in false areas. The false candidates have to be filtered out with
further techniques, described in the next section.

3 Object Contour Detection with Saliency Functions and

Graph Theory

3.1 Detection of Local Structures

According to [16], local contours around keypoints are ecient, low dimensional
tool for matching and distinguishing, therefore this algorithm was now imple-
mented for Harris keypoint candidates to filter out the falsely detected points.
The main steps for estimating local structure characteristics:

1. Generating Harris keypoints for difference map (Section 2.2)
2. Generating the Local Contour around keypoints in the original image [17]
3. Calculating Modified Fourier Descriptor for the estimated closed curve [18]
4. Describe the contour by a limited set of Fourier coecients [19]

As the specification shows, after detecting the Harris keypoint candidates (the
method is briefly summarized in Section 2.2), GVF Snake [17] was used for local
contour ( LC ) analysis. LC was computed in the original image, in a 20

20 size
area, where the keypoint was in the middle. The generated LC assigns an individ-
ual shape to every keypoint, but the dimension is quite high. Therefore modified
Fourier descriptors were applied, which represents the LC at low dimension. We
have determined the cut-off frequency by maximizing the recognition accuracies
and minimizing the noise of irregularities of the shape boundaries and chose the
first twenty coecients (excluding the DC component to remove the positional
sensitivity).

×

3.2 Matching with Local Contours

Our assumption was that after having the FDs for the keypoints, differences
between keypoint surroundings can be searched through this descriptor set. We
extended the MFD method to get symmetric distance computation as it is writ-
ten in [16]. By comparing a keypoint ( p i ) on the first frame and on the second
frame, D i represents the similarity value. If the following criteria exists:

D i > 3 (4)

where 3 = 3 is a tolerance value, than the keypoint is supposed to be a changed
area.
Results of the detection is provided in Figure 5(a).