Information Technology Reference
In-Depth Information
analyzer modules [1], [2]. While numerous state-of-the art approaches in remote
sensing deal with multispectral [3], [4], [5], [6] or synthetic aperture radar (SAR)
[7], [8] imagery, the significance of handling optical photographs is also increas-
ing [9]. Here, the processing methods should consider that several optical image
collections include partially archive data, where the photographs are grayscale
or contain only poor color information. This paper focuses on finding contours
of newly appearing/fading out objects in optical aerial images which were taken
with several years time differences partially in different seasons and in different
lighting conditions. In this case, simple techniques like thresholding the differ-
ence image [10] or background modeling [11] cannot be adopted eciently since
details are not comparable.
These optical image sensors provide limited information and we can only
assume to have image repositories which contain geometrically corrected and
registered [12] grayscale orthophotographs.
In the literature one main group of approaches is the postclassification com-
parison, which segments the input images with different land-cover classes, like
arboreous lands, barren lands, and artificial structures [13], obtaining the changes
indirectly as regions with different classes in the two image layers [9]. We follow
another methodology, like direct methods [3], [5], [5], [7], where a similarity-
feature map from the input photographs (e.g., a DI) is derived, then the feature
map is separated into changed and unchanged areas.
Our direct method does not use any land-cover class models, and attempts
to detect changes which can be discriminated by low-level features. However,
our approach is not a pixel-neighborhood MAP system as in [14], but a connec-
tion system of nearby saliency points. These saliency points define a connection
system by using local graphs for outlining the local hull of the objects. Consid-
ering this curve as a starting spline, we search for objects' boundaries by active
contour iterations.
The above features are local saliency points and saliency functions. The main
saliency detector is calculated as a discrimative function among the functions
of the different layers. We show that Harris detector is the appropriate function
for finding the dissimilarities among different layers, when comparison is not
possible because of the different lighting, color and contrast conditions.
Local structure around keypoints is investigated by generating scale and po-
sition invariant descriptors, like SIFT. These descriptors describe the local mi-
crostructure, however, in several cases more succinct set of parameters is needed.
For this reason we have developed a local active-contour based descriptor around
keypoints, but this contour is generated by edginess in the cost function, while
we characterize keypoints of junctions. To fit together the definition of keypoints
and their active contour around them, we have introduced the Harris corner de-
tection as an outline detector instead of the simple edge functions. This change
resulted in a much better characterization of local structure.
In the following, we introduce a new change detection procedure by using
Harris function and its derivatives for finding saliency points among image sam-
ples; then a new local descriptor will be demonstated by generating local active
Search WWH ::
Custom Search