def susanCorner(image, radius=3.4, fptype=bool, t=25, gfrac=None,
cgthresh=None, nshades=256):
"""
Implements SUSAN corner detector algorithm as described in "SUSAN--A
New Approach to Low Level Image Processing", S. M. Smith and J. M.
Brady, Technical Report TR95SMSIc (1995), available at:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.2763 .
Alternatively, there is also a slightly abridged version of the
same reference, with identical title and authors, available at
International Journal of Computer Vision 23(1): 45-78 (1997).
image: array_like
Array containing grayscale image data. Only grayscale pixel
values are supported--cannot handle 3-channel color.
radius: scalar, optional
Circular footprint radius, in units of pixels; passed down
as input to masks.circFootprint(). Default is 3.4, as
recommended in reference.
fptype: data-type, optional
Data type of circular footprint; passed to masks.circFootprint().
Default is bool, since the version of the algorithm originally
described in the reference did not define the algorithm behavior
for any other type of footprint (we have trivially extended it
to cover float type footprints as well, however there is not
much observable different in the results, as it only affects the
weighting of the pixels at the edge of the footprint, not the
central region).
t: scalar, optional
Threshold value for color difference required to exclude/include
a pixel in the USAN region described in the reference article.
Default is 25 grayscale levels, as suggested by article authors,
and assumes a grayscale image with a total of 256 distinct
shades. As described in eq. 4, this is technically a "soft"
threshold, rather than a hard cutoff, and moreover it's also
bidirectional; i.e., a setting of t=25 actually means +/- 25.
gfrac: scalar, optional
Fraction of maximum number of USAN pixels to assign to the "g"
parameter in eq. 3 of the reference article. Default setting
of gfrac=None triggers lower level functions to select a
context-appropriate default value (for corner detection
specifically, default is gfrac=0.5) as recommended by
reference article authors.
cgthesh: scaler, optional
Threshold value of USAN center of gravity, in units of pixels,
which will lead to a shift between one underlying set of
approximations/assumptions vs. another. Default setting of
cgthresh=None triggers lower level masks.usan() function to
select a context-appropriate default value (for corner detection
specifically, default is cgthresh=0.45*radius) as recommended
by reference article authors. See masks.usan() source code
and reference article for further (extremely esoteric) details.
For corner detection in particular (less so with edge detection)
we recommend testing a variety of settings both above and below
the default as this parameter seems to have a particularly
strong effect in some cases on signal vs. noise.
nshades: scalar, optional
Total number of distinct integer grayscale levels available in
image format. Defaults to 256 (i.e., 2**8), appropriate for an
8 bit image format such as jpeg. For image formats with higher
grayscale color resolution, e.g. such as 12-bit or 16-bit, set
nshades=4096 or nshades=65536 (i.e., 2**12 or 2**16).
Returns
-------
corner: ndarray
The corner response of the image as calculated by the SUSAN corner
detection algorithm.
"""
# Get raw corner response; many corners may consist of small clusters of
# adjacent responsive pixels
rawcorner = masks.usan(image, mode='Corner', radius=radius,
fptype=fptype, t=t, gfrac=gfrac,
cgthresh=cgthresh, nshades=nshades)
# Find maximum corner response within circular USAN footprint (but force
# footprint type to be bool in this case regardless of user-selected
# input fptype, because float would make no sense in this context)
fp = masks.circFootprint(radius=radius, dtype=bool)
rawmax = filters.maximum_filter(rawcorner, footprint=fp)
# True corners are those where response is both locally maximum as well
# as non-zero
corner = np.where(rawcorner == rawmax, rawcorner, 0)
return corner
def susanCorner(image,
radius=3.4,
fptype=bool,
t=25,
gfrac=None,
cgthresh=None,
nshades=256):
"""
Implements SUSAN corner detector algorithm as described in "SUSAN--A
New Approach to Low Level Image Processing", S. M. Smith and J. M.
Brady, Technical Report TR95SMSIc (1995), available at:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.2763 .
Alternatively, there is also a slightly abridged version of the
same reference, with identical title and authors, available at
International Journal of Computer Vision 23(1): 45-78 (1997).
image: array_like
Array containing grayscale image data. Only grayscale pixel
values are supported--cannot handle 3-channel color.
radius: scalar, optional
Circular footprint radius, in units of pixels; passed down
as input to masks.circFootprint(). Default is 3.4, as
recommended in reference.
fptype: data-type, optional
Data type of circular footprint; passed to masks.circFootprint().
Default is bool, since the version of the algorithm originally
described in the reference did not define the algorithm behavior
for any other type of footprint (we have trivially extended it
to cover float type footprints as well, however there is not
much observable different in the results, as it only affects the
weighting of the pixels at the edge of the footprint, not the
central region).
t: scalar, optional
Threshold value for color difference required to exclude/include
a pixel in the USAN region described in the reference article.
Default is 25 grayscale levels, as suggested by article authors,
and assumes a grayscale image with a total of 256 distinct
shades. As described in eq. 4, this is technically a "soft"
threshold, rather than a hard cutoff, and moreover it's also
bidirectional; i.e., a setting of t=25 actually means +/- 25.
gfrac: scalar, optional
Fraction of maximum number of USAN pixels to assign to the "g"
parameter in eq. 3 of the reference article. Default setting
of gfrac=None triggers lower level functions to select a
context-appropriate default value (for corner detection
specifically, default is gfrac=0.5) as recommended by
reference article authors.
cgthesh: scaler, optional
Threshold value of USAN center of gravity, in units of pixels,
which will lead to a shift between one underlying set of
approximations/assumptions vs. another. Default setting of
cgthresh=None triggers lower level masks.usan() function to
select a context-appropriate default value (for corner detection
specifically, default is cgthresh=0.45*radius) as recommended
by reference article authors. See masks.usan() source code
and reference article for further (extremely esoteric) details.
For corner detection in particular (less so with edge detection)
we recommend testing a variety of settings both above and below
the default as this parameter seems to have a particularly
strong effect in some cases on signal vs. noise.
nshades: scalar, optional
Total number of distinct integer grayscale levels available in
image format. Defaults to 256 (i.e., 2**8), appropriate for an
8 bit image format such as jpeg. For image formats with higher
grayscale color resolution, e.g. such as 12-bit or 16-bit, set
nshades=4096 or nshades=65536 (i.e., 2**12 or 2**16).
Returns
-------
corner: ndarray
The corner response of the image as calculated by the SUSAN corner
detection algorithm.
"""
# Get raw corner response; many corners may consist of small clusters of
# adjacent responsive pixels
rawcorner = masks.usan(image,
mode='Corner',
radius=radius,
fptype=fptype,
t=t,
gfrac=gfrac,
cgthresh=cgthresh,
nshades=nshades)
# Find maximum corner response within circular USAN footprint (but force
# footprint type to be bool in this case regardless of user-selected
# input fptype, because float would make no sense in this context)
fp = masks.circFootprint(radius=radius, dtype=bool)
rawmax = filters.maximum_filter(rawcorner, footprint=fp)
# True corners are those where response is both locally maximum as well
# as non-zero
corner = np.where(rawcorner == rawmax, rawcorner, 0)
return corner
def susanEdge(image, radius=3.4, fptype=bool, t=25, gfrac=None, cgthresh=None,
nshades=256, tlo=0.1, thi=0.3, minmem=1):
"""
Implements SUSAN edge detector algorithm as described in "SUSAN--A
New Approach to Low Level Image Processing", S. M. Smith and J. M.
Brady, Technical Report TR95SMSIc (1995), available at:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.2763 .
Alternatively, there is also a slightly abridged version of the
same reference, with identical title and authors, available at
International Journal of Computer Vision 23(1): 45-78 (1997).
image: array_like
Array containing grayscale image data. Only grayscale pixel
values are supported--cannot handle 3-channel color.
radius: scalar, optional
Circular footprint radius, in units of pixels; passed down
as input to masks.circFootprint(). Default is 3.4, as
recommended in reference.
fptype: data-type, optional
Data type of circular footprint; passed to masks.circFootprint().
Default is bool, since the version of the algorithm originally
described in the reference did not define the algorithm behavior
for any other type of footprint (we have trivially extended it
to cover float type footprints as well, however there is not
much observable different in the results, as it only affects the
weighting of the pixels at the edge of the footprint, not the
central region).
t: scalar, optional
Threshold value for color difference required to exclude/include
a pixel in the USAN region described in the reference article.
Default is 25 grayscale levels, as suggested by article authors,
and assumes a grayscale image with a total of 256 distinct
shades. As described in eq. 4, this is technically a "soft"
threshold, rather than a hard cutoff, and moreover it's also
bidirectional; i.e., a setting of t=25 actually means +/- 25.
gfrac: scalar, optional
Fraction of maximum number of USAN pixels to assign to the "g"
parameter in eq. 3 of the reference article. Default setting
of gfrac=None triggers lower level masks.usan() function to
select a context-appropriate default value (for edge detection
specifically, default is gfrac=0.75) as recommended by
reference article authors.
cgthesh: scaler, optional
Threshold value of USAN center of gravity, in units of pixels,
which will lead to a shift between one underlying set of
approximations/assumptions vs. another. Default setting of
cgthresh=None triggers lower level masks.usan() function to
select a context-appropriate default value (for edge detection
specifically, default is cgthresh=0.3*radius) as recommended
by reference article authors. See masks.usan() source code
and reference article for further (extremely esoteric) details.
nshades: scalar, optional
Total number of distinct integer grayscale levels available in
image format. Defaults to 256 (i.e., 2**8), appropriate for an
8 bit image format such as jpeg. For image formats with higher
grayscale color resolution, e.g. such as 12-bit or 16-bit, set
nshades=4096 or nshades=65536 (i.e., 2**12 or 2**16).
tlo, thi: scalar, optional
Low and high thresholds for Canny-style hysteresis thresholding.
Default values are 0.1 and 0.3, and represent the two thresholds
as fractions of the maximum edge response for the entire image.
Set tlo=thi in order to turn off dual thresholding and revert
effectively to a single threshold scheme. Set tlo=thi=0 to turn
off thresholding altogether. All values must lie on interval
0 <= tlo <= thi <= 1.
minmem: scalar, optional
Specifies minimum pixel membership to be required in each valid
edge. E.g., minmem=4 will delete extremely short junk edges
consisting of only 1, 2, or 3 adjacent pixels, but will keep
edges with 4 or more pixels. This is sometimes useful when
processing an image with texture, e.g., such as a brick wall,
where small segments of the texture itself, such as the mortar
between the bricks, may occasionally be misinterpreted by the
algorithm as extremely short macroscopic edges. Default is
set to minmem=1, which effectively leaves this section of the
processing turned off in normal usage, due to the fact that it
was not part of the SUSAN algorithm as originally defined by
the reference article authors.
Returns
-------
edge: ndarray
The edge response of the image as calculated by the SUSAN edge
detection algorithm.
"""
# Get raw edge response; many edges may be wider than one pixel
thickedge = masks.usan(image, mode='Edge', radius=radius, fptype=fptype,
t=t, gfrac=gfrac, nshades=nshades)
# Get angle (in degrees) of normal vector perpendicular to each edge
edgedir = masks.usan(image, mode='EdgeDir', radius=radius, fptype=fptype,
t=t, gfrac=gfrac, cgthresh=cgthresh, nshades=nshades)
# Round normal vector to one of four cardinal directions: left-right,
# up-down, diagonal (positive slope), diagonal (negative slope)
rounddir = cleanedges.roundAngle(edgedir)
# Thin edges to single pixel width by accepting only those whose USAN
# response is at a local maximum relative to the local edge gradient
maxedge = cleanedges.nonMaxSuppEdge(thickedge, rounddir)
# Apply Canny-style hysteresis thresholding
htedge = cleanedges.hystThresh(maxedge, tlo, thi)
# Delete edges consisting of fewer than minmem connected pixels
edge = cleanedges.minMembership(htedge, minmem=minmem)
return edge
def susanEdge(image,
radius=3.4,
fptype=bool,
t=25,
gfrac=None,
cgthresh=None,
nshades=256,
tlo=0.1,
thi=0.3,
minmem=1):
"""
Implements SUSAN edge detector algorithm as described in "SUSAN--A
New Approach to Low Level Image Processing", S. M. Smith and J. M.
Brady, Technical Report TR95SMSIc (1995), available at:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.2763 .
Alternatively, there is also a slightly abridged version of the
same reference, with identical title and authors, available at
International Journal of Computer Vision 23(1): 45-78 (1997).
image: array_like
Array containing grayscale image data. Only grayscale pixel
values are supported--cannot handle 3-channel color.
radius: scalar, optional
Circular footprint radius, in units of pixels; passed down
as input to masks.circFootprint(). Default is 3.4, as
recommended in reference.
fptype: data-type, optional
Data type of circular footprint; passed to masks.circFootprint().
Default is bool, since the version of the algorithm originally
described in the reference did not define the algorithm behavior
for any other type of footprint (we have trivially extended it
to cover float type footprints as well, however there is not
much observable different in the results, as it only affects the
weighting of the pixels at the edge of the footprint, not the
central region).
t: scalar, optional
Threshold value for color difference required to exclude/include
a pixel in the USAN region described in the reference article.
Default is 25 grayscale levels, as suggested by article authors,
and assumes a grayscale image with a total of 256 distinct
shades. As described in eq. 4, this is technically a "soft"
threshold, rather than a hard cutoff, and moreover it's also
bidirectional; i.e., a setting of t=25 actually means +/- 25.
gfrac: scalar, optional
Fraction of maximum number of USAN pixels to assign to the "g"
parameter in eq. 3 of the reference article. Default setting
of gfrac=None triggers lower level masks.usan() function to
select a context-appropriate default value (for edge detection
specifically, default is gfrac=0.75) as recommended by
reference article authors.
cgthesh: scaler, optional
Threshold value of USAN center of gravity, in units of pixels,
which will lead to a shift between one underlying set of
approximations/assumptions vs. another. Default setting of
cgthresh=None triggers lower level masks.usan() function to
select a context-appropriate default value (for edge detection
specifically, default is cgthresh=0.3*radius) as recommended
by reference article authors. See masks.usan() source code
and reference article for further (extremely esoteric) details.
nshades: scalar, optional
Total number of distinct integer grayscale levels available in
image format. Defaults to 256 (i.e., 2**8), appropriate for an
8 bit image format such as jpeg. For image formats with higher
grayscale color resolution, e.g. such as 12-bit or 16-bit, set
nshades=4096 or nshades=65536 (i.e., 2**12 or 2**16).
tlo, thi: scalar, optional
Low and high thresholds for Canny-style hysteresis thresholding.
Default values are 0.1 and 0.3, and represent the two thresholds
as fractions of the maximum edge response for the entire image.
Set tlo=thi in order to turn off dual thresholding and revert
effectively to a single threshold scheme. Set tlo=thi=0 to turn
off thresholding altogether. All values must lie on interval
0 <= tlo <= thi <= 1.
minmem: scalar, optional
Specifies minimum pixel membership to be required in each valid
edge. E.g., minmem=4 will delete extremely short junk edges
consisting of only 1, 2, or 3 adjacent pixels, but will keep
edges with 4 or more pixels. This is sometimes useful when
processing an image with texture, e.g., such as a brick wall,
where small segments of the texture itself, such as the mortar
between the bricks, may occasionally be misinterpreted by the
algorithm as extremely short macroscopic edges. Default is
set to minmem=1, which effectively leaves this section of the
processing turned off in normal usage, due to the fact that it
was not part of the SUSAN algorithm as originally defined by
the reference article authors.
Returns
-------
edge: ndarray
The edge response of the image as calculated by the SUSAN edge
detection algorithm.
"""
# Get raw edge response; many edges may be wider than one pixel
thickedge = masks.usan(image,
mode='Edge',
radius=radius,
fptype=fptype,
t=t,
gfrac=gfrac,
nshades=nshades)
# Get angle (in degrees) of normal vector perpendicular to each edge
edgedir = masks.usan(image,
mode='EdgeDir',
radius=radius,
fptype=fptype,
t=t,
gfrac=gfrac,
cgthresh=cgthresh,
nshades=nshades)
# Round normal vector to one of four cardinal directions: left-right,
# up-down, diagonal (positive slope), diagonal (negative slope)
rounddir = cleanedges.roundAngle(edgedir)
# Thin edges to single pixel width by accepting only those whose USAN
# response is at a local maximum relative to the local edge gradient
maxedge = cleanedges.nonMaxSuppEdge(thickedge, rounddir)
# Apply Canny-style hysteresis thresholding
htedge = cleanedges.hystThresh(maxedge, tlo, thi)
# Delete edges consisting of fewer than minmem connected pixels
edge = cleanedges.minMembership(htedge, minmem=minmem)
return edge
