Automatic Recognition of Impact Craters on the Martian Surface from DEM and Images - Planetary Exploration and Science: Recent Results and Advances - page 107

Fig. 6.4 The left is the binary image generated by the flooded algorithm, and the right one is the
segmentation image by bounding box
as a stand-alone crater-detection algorithm. However, it can be used as assistant
approach to reduce the computational cost by segmentation of the site image.
It will produce numerous fragments after the segmentation of the image into
connected components, which are labeled by i D 1, 2, :::, k . Each fragment contains
some degree of topographic depression, but not all of them are craters and some are
even not craters. For each fragment, we calculate a bounding box. To allow for
situations where craters can be slightly larger than fragments due to low top off
point caused by rim degradation, each of them is slightly larger than the extent of its
enclosed fragment. Because I f and I k are co-registered images, applying bounding
boxes B i to I k divides a binary image of threshold profile curvature into small
images, without cutting through the craters. We detect craters from each small image
separately. This procedure guarantees that all detectable craters in the entire site are
detected (possibly more than once). The benefit of site segmentation is the much-
reduced computational cost of crater detection (Fig. 6.4 ).
Selecting Crater Candidate
After the first step in processing DEM data, we can easily identify the crater rims
by using the binary profile curvature map. There are a few more steps before we
can detect craters in the binary images. First, the morphological closing operation
followed by a thinning operation needs to be done with a 3 3 structuring element to
smooth the edges of structures and to eliminate small holes in them. Then, in order
to reduce all lines to a single pixel thickness, we utilize the thinning operation for
skeletonization. In the next step, we can use the circular Hough transform to detect
crater candidates (Bue and Stepinski 2007 ). In the third step, we will examine each
candidate using our confirmation algorithm which either accepts a candidate as a
crater and adds it to the catalogue or rejects it. Finally, an elimination algorithm will
be run on the entire catalogue to remove duplicate detections.