Binarization results of poor quality gray-level images does
not guarantee a consistent number neither the quality of resulted components, therefore, poses real limitation for using
the Shape Context descriptor. The presented descriptor uses
view points taken from rings out of the shape and therefore
not significantly affected by the results of the binarization
step. In such cases, limited results of the binarization process can still serve to determine the values of diameter (D),
height H, and centroid C as in the binary case. From the
other hand, since the contour results are not guaranteed,
we modify the presented descriptor to captures the weighted
distribution over relative positions of the shape pixels. Multi
resolution rings around the centroid can still be used for
richer descriptions. In this case each feature descriptor from
a view point is a log-polar weighted histogram of the coordinates of each pixel in the shape. The gray level value
of pixels are used as weights, and the reference points are
the view points. In poor quality images, one may emphasize
edges using Sobel or Gaussian low pass edge-emphasizing
filters for better performance. In our case we have used the
Sobel edge-emphasizing filter on the gray-evel image. A predefined low value have been used to binarize only the back
ground of the image and later on calculate the values C and
D for the centroid and Diameter respectively. The ring centered on C with the diameter D have been used to generate
n view points taken uniformly distant on the ring.
Binarization results of poor quality gray-level images does
not guarantee a consistent number neither the quality of resulted components, therefore, poses real limitation for using
the Shape Context descriptor. The presented descriptor uses
view points taken from rings out of the shape and therefore
not significantly affected by the results of the binarization
step. In such cases, limited results of the binarization process can still serve to determine the values of diameter (D),
height H, and centroid C as in the binary case. From the
other hand, since the contour results are not guaranteed,
we modify the presented descriptor to captures the weighted
distribution over relative positions of the shape pixels. Multi
resolution rings around the centroid can still be used for
richer descriptions. In this case each feature descriptor from
a view point is a log-polar weighted histogram of the coordinates of each pixel in the shape. The gray level value
of pixels are used as weights, and the reference points are
the view points. In poor quality images, one may emphasize
edges using Sobel or Gaussian low pass edge-emphasizing
filters for better performance. In our case we have used the
Sobel edge-emphasizing filter on the gray-evel image. A predefined low value have been used to binarize only the back
ground of the image and later on calculate the values C and
D for the centroid and Diameter respectively. The ring centered on C with the diameter D have been used to generate
n view points taken uniformly distant on the ring.
การแปล กรุณารอสักครู่..