Top |
Functions
int | vips_merge () |
int | vips_mosaic () |
int | vips_mosaic1 () |
int | vips_match () |
int | vips_globalbalance () |
int | vips_remosaic () |
int | im_correl () |
int | im_align_bands () |
int | im_maxpos_subpel () |
Description
These functions are useful for joining many small images together to make one large image. They can cope with unstable contrast and arbitary sub-image layout, but will not do any geometric correction. Geometric errors should be removed before using these functions.
The mosaicing functions can be grouped into layers:
The lowest level functions are im_correl()
and vips_merge()
.
im_correl()
searches a large image for a small sub-image, returning
the position of the best sub-image match. vips_merge()
joins two images together
left-right or up-down with a smooth seam.
Next, vips_mosaic()
use the
search function plus the two low-level merge operations to join two images
given just an approximate overlap as a start point.
The functions im_lrmosaic1()
and im_tbmosaic1()
are
first-order
analogues of the basic mosaic functions: they take two approximate
tie-points and use
them to rotate and scale the right-hand or bottom image before starting to
join.
Finally, vips_globalbalance()
can be used to remove contrast differences in
a mosaic
which has been assembled with these functions. It takes the mosaic apart,
measures image contrast differences along the seams, finds a set of
correction factors which will minimise these differences, and reassembles
the mosaic.
im_remosaic()
uses the
same
techniques, but will reassemble the image from a different set of source
images.
Functions
vips_merge ()
int vips_merge (VipsImage *ref
,VipsImage *sec
,VipsImage **out
,VipsDirection direction
,int dx
,int dy
,...
);
Optional arguments:
mblend
: maximum blend size
This operation joins two images left-right (with ref
on the left) or
up-down (with ref
above) with a smooth seam.
If the number of bands differs, one of the images must have one band. In this case, an n-band image is formed from the one-band image by joining n copies of the one-band image together, and then the two n-band images are operated upon.
The two input images are cast up to the smallest common type (see table Smallest common format in
arithmetic).dx
and dy
give the displacement of sec
relative to ref
, in other words,
the vector to get from the origin of sec
to the origin of ref
, in other
words, dx
will generally be a negative number.
mblend
limits the maximum width of the
blend area. A value of "-1" means "unlimited". The two images are blended
with a raised cosine.
Pixels with all bands equal to zero are "transparent", that is, zero pixels in the overlap area do not contribute to the merge. This makes it possible to join non-rectangular images.
See also: vips_mosaic()
, vips_insert()
.
Parameters
ref |
reference image |
|
sec |
secondary image |
|
out |
output image |
|
direction |
horizontal or vertical merge |
|
dx |
displacement of ref from sec |
|
dy |
displacement of ref from sec |
|
... |
|
vips_mosaic ()
int vips_mosaic (VipsImage *ref
,VipsImage *sec
,VipsImage **out
,VipsDirection direction
,int xref
,int yref
,int xsec
,int ysec
,...
);
Optional arguments:
bandno
: band to search for features
hwindow
: half window size
harea
: half search size
mblend
: maximum blend size
This operation joins two images left-right (with ref
on the left) or
top-bottom (with ref
above) given an approximate overlap.
sec
is positioned so that the pixel (xsec
, ysec
) lies on top of the
pixel in ref
at (xref
, yref
). The overlap area is divided into three
sections, 20 high-contrast points in band bandno
of image ref
are found
in each, and each high-contrast point is searched for in sec
using
hwindow
and harea
(see vips_correl()
).
A linear model is fitted to the 60 tie-points, points a long way from the fit are discarded, and the model refitted until either too few points remain or the model reaches good agreement.
The detected displacement is used with vips_merge()
to join the two images
together.
See also: vips_merge()
, vips_insert()
.
vips_mosaic1 ()
int vips_mosaic1 (VipsImage *ref
,VipsImage *sec
,VipsImage **out
,VipsDirection direction
,int xr1
,int yr1
,int xs1
,int ys1
,int xr2
,int yr2
,int xs2
,int ys2
,...
);
Optional arguments:
search
: search to improve tie-points
hwindow
: half window size
harea
: half search size
interpolate
: interpolate pixels with this
mblend
: maximum blend size
bandno
: band to search for features
This operation joins two images top-bottom (with sec
on the right)
or left-right (with sec
at the bottom)
given an approximate pair of tie-points. sec
is scaled and rotated as
necessary before the join.
Before performing the transformation, the tie-points are improved by
searching band bandno
in an area of sec
of size hsearchsize
for a
match of size hwindowsize
to ref
.
If search
is TRUE
, before performing the transformation, the tie-points
are improved by searching an area of sec
of size harea
for a
mosaic1 of size hwindow
to ref
.
mblend
limits the maximum size of the
blend area. A value of "-1" means "unlimited". The two images are blended
with a raised cosine.
Pixels with all bands equal to zero are "transparent", that is, zero pixels in the overlap area do not contribute to the merge. This makes it possible to join non-rectangular images.
If the number of bands differs, one of the images must have one band. In this case, an n-band image is formed from the one-band image by joining n copies of the one-band image together, and then the two n-band images are operated upon.
The two input images are cast up to the smallest common type (see table Smallest common format in
arithmetic).See also: vips_merge()
, vips_insert()
, vips_globalbalance()
.
Parameters
ref |
reference image |
|
sec |
secondary image |
|
out |
output image |
|
direction |
horizontal or vertical join |
|
xr1 |
first reference tie-point |
|
yr1 |
first reference tie-point |
|
xs1 |
first secondary tie-point |
|
ys1 |
first secondary tie-point |
|
xr2 |
second reference tie-point |
|
yr2 |
second reference tie-point |
|
xs2 |
second secondary tie-point |
|
ys2 |
second secondary tie-point |
vips_match ()
int vips_match (VipsImage *ref
,VipsImage *sec
,VipsImage **out
,int xr1
,int yr1
,int xs1
,int ys1
,int xr2
,int yr2
,int xs2
,int ys2
,...
);
Optional arguments:
search
: search to improve tie-points
hwindow
: half window size
harea
: half search size
interpolate
: interpolate pixels with this
Scale, rotate and translate sec
so that the tie-points line up.
If search
is TRUE
, before performing the transformation, the tie-points
are improved by searching an area of sec
of size harea
for a
match of size hwindow
to ref
.
This function will only work well for small rotates and scales.
Parameters
ref |
reference image |
|
sec |
secondary image |
|
out |
output image |
|
xr1 |
first reference tie-point |
|
yr1 |
first reference tie-point |
|
xs1 |
first secondary tie-point |
|
ys1 |
first secondary tie-point |
|
xr2 |
second reference tie-point |
|
yr2 |
second reference tie-point |
|
xs2 |
second secondary tie-point |
|
ys2 |
second secondary tie-point |
vips_globalbalance ()
int vips_globalbalance (VipsImage *in
,VipsImage **out
,...
);
Optional arguments:
gamma
: gamma of source images
int_output
: TRUE
for integer image output
vips_globalbalance() can be used to remove contrast differences in an assembled mosaic.
It reads the History field attached to in
and builds a list of the source
images that were used to make the mosaic and the position that each ended
up at in the final image.
It opens each of the source images in turn and extracts all parts which
overlap with any of the other images. It finds the average values in the
overlap areas and uses least-mean-square to find a set of correction
factors which will minimise overlap differences. It uses gamma
to
gamma-correct the source images before calculating the factors. A value of
1.0 will stop this.
Each of the source images is transformed with the appropriate correction
factor, then the mosaic is reassembled. out
is VIPS_FORMAT_FLOAT, but
if int_output
is set, the output image is the same format as the input
images.
There are some conditions that must be met before this operation can work:
the source images must all be present under the filenames recorded in the
history on in
, and the mosaic must have been built using only operations in
this package.
See also: vips_remosaic()
.
Parameters
in |
mosaic to rebuild |
|
out |
output image |
|
... |
|
vips_remosaic ()
int vips_remosaic (VipsImage *in
,VipsImage **out
,const char *old_str
,const char *new_str
,...
);
vips_remosaic() works rather as vips_globalbalance()
. It takes apart the
mosaiced image in
and rebuilds it, substituting images.
Unlike vips_globalbalance()
, images are substituted based on their file‐
names. The rightmost occurence of the string old_str
is swapped
for new_str
, that file is opened, and that image substituted for
the old image.
It's convenient for multispectral images. You can mosaic one band, then use that mosaic as a template for mosaicing the others automatically.
See also: vips_globalbalance()
.
Parameters
in |
mosaic to rebuild |
|
out |
output image |
|
old_str |
gamma of source images |
|
new_str |
gamma of source images |
|
... |
|
im_correl ()
int im_correl (VipsImage *ref
,VipsImage *sec
,int xref
,int yref
,int xsec
,int ysec
,int hwindowsize
,int hsearchsize
,double *correlation
,int *x
,int *y
);
This operation finds the position of sec
within ref
.
The area around
(xsec
, ysec
) is searched for the best match to the area around (xref
,
yref
). It searches an area of size hsearchsize
for a
match of size hwindowsize
. The position of the best match is
returned, together with the correlation at that point.
Only the first band of each image is correlated. ref
and sec
may be
very large --- the function extracts and generates just the
parts needed. Correlation is done with im_spcor()
; the position of
the maximum is found with im_maxpos()
.
See also: im_match_linear()
, im_match_linear_search()
, im_lrmosaic()
.
Parameters
ref |
reference image |
|
sec |
secondary image |
|
xref |
position in reference image |
|
yref |
position in reference image |
|
xsec |
position in secondary image |
|
ysec |
position in secondary image |
|
hwindowsize |
half window size |
|
hsearchsize |
half search size |
|
correlation |
return detected correlation |
|
x |
return found position |
|
y |
return found position |
im_align_bands ()
int im_align_bands (VipsImage *in
,VipsImage *out
);
This operation uses im_phasecor_fft()
to find an integer displacement to
align all image bands band 0. It is very slow and not very accurate.
Use im_estpar()
in preference: it's fast and accurate.
See also: im_global_balancef()
, im_remosaic()
.
im_maxpos_subpel ()
int im_maxpos_subpel (VipsImage *in
,double *x
,double *y
);
This function implements:
"Extension of Phase Correlation to Subpixel Registration" by H. Foroosh, from IEEE trans. Im. Proc. 11(3), 2002.
If the best three matches in the correlation are aranged:
02 or 01 1 2
then we return a subpixel match using the ratio of correlations in the vertical and horizontal dimension.
( xs[0], ys[0] ) is the best integer alignment ( xs[ use_x ], ys[ use_x ] ) is equal in y and (+/-)1 off in x ( xs[ use_y ], ys[ use_y ] ) is equal in x and (+/-)1 off in y
Alternatively if the best four matches in the correlation are aranged in a square:
01 or 03 or 02 or 03 32 12 31 21
then we return a subpixel match weighting with the sum the two on each side over the sum of all four, but only if all four of them are very close to the best, and the fifth is nowhere near.
This alternative method is not described by Foroosh, but is often the case where the match is close to n-and-a-half pixels in both dimensions.
See also: im_maxpos()
, im_min()
, im_stats()
.