Orientation selectivity
Ranklets share the same orientation
selectivity pattern as Haar wavelets.
This is intuitively evident if we remember that a ranklet essentially counts in how many of the pairs in TxC the pixel from the T set is brighter than the one from the C set. So, for example,
This is intuitively evident if we remember that a ranklet essentially counts in how many of the pairs in TxC the pixel from the T set is brighter than the one from the C set. So, for example,
- This ranklet straddles a vertical dark-to-bright edge. Most
pixels in the T
set are
brighter than those in the C
set, which will result in a large count (positive
edge response)
- If the edge is inverted, most pixels in the C set are
brighter
than those in the T set, which
results in a low count
(negative edge response)
- In this case, the edge does not match the orientation tuning of
this ranklet. Pixel brightness is uniformly distributed across the
T and C sets, leading to an average count
(null response)
Thus, this ranklet responds to vertical structures in the image. In the same way, the other two ranklets respond to horizontal edges and diagonal lines (or corners) respectively.
With an appropriate normalization, it is easy to map the responses to the [-1,+1] range, so that a null response will in fact correspond to a zero.
| www.Ranklets.net | F. Smeraldi, June 2004 |