libvips/libsrc/mosaicing/yafrtest.cpp

699 lines
21 KiB
C++

/* vipsinterpolateyafr_test ... yarf as a vips interpolate class
*/
/*
This file is part of VIPS.
VIPS is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
/*
These files are distributed with VIPS - http://www.vips.ecs.soton.ac.uk
*/
/*
#define DEBUG
*/
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif /*HAVE_CONFIG_H*/
#include <vips/intl.h>
#include <stdio.h>
#include <stdlib.h>
#include <vips/vips.h>
#include <vips/internal.h>
#ifdef WITH_DMALLOC
#include <dmalloc.h>
#endif /*WITH_DMALLOC*/
/* "fast" floor() ... on my laptop, anyway.
*/
#define FLOOR( V ) ((V) >= 0 ? (int)(V) : (int)((V) - 1))
#ifndef restrict
#ifdef __restrict
#define restrict __restrict
#else
#ifdef __restrict__
#define restrict __restrict__
#else
#define restrict
#endif
#endif
#endif
static VipsInterpolateClass *vips_interpolate_yafr_test_parent_class = NULL;
/*
* 2008 (c) Nicolas Robidoux (developer of Yet Another Fast
* Resampler).
*
* Acknowledgement: N. Robidoux's research on YAFR_TEST funded in part by
* an NSERC (National Science and Engineering Research Council of
* Canada) Discovery Grant.
*/
/*
* YAFR_TEST = Yet Another Fast Resampler
*
* Yet Another Fast Resampler is a nonlinear resampler which consists
* of a linear scheme (in this version, Catmull-Rom) plus a nonlinear
* sharpening correction the purpose of which is the straightening of
* diagonal interfaces between flat colour areas.
*
* Key properties:
*
* YAFR_TEST (smooth) is interpolatory:
*
* If asked for the value at the center of an input pixel, it will
* return the corresponding value, unchanged.
*
* YAFR_TEST (smooth) preserves local averages:
*
* The average of the reconstructed intensity surface over any region
* is the same as the average of the piecewise constant surface with
* values over pixel areas equal to the input pixel values (the
* "nearest neighbour" surface), except for a small amount of blur at
* the boundary of the region. More precicely: YAFR_TEST (smooth) is a box
* filtered exact area method.
*
* Main weaknesses of YAFR_TEST (smooth):
*
* Weakness 1: YAFR_TEST (smooth) improves on Catmull-Rom only for images
* with at least a little bit of smoothness.
*
* Weakness 2: Catmull-Rom introduces a lot of haloing. YAFR_TEST (smooth)
* is based on Catmull-Rom, and consequently it too introduces a lot
* of haloing.
*
* More details regarding Weakness 1:
*
* If a portion of the image is such that every pixel has immediate
* neighbours in the horizontal and vertical directions which have
* exactly the same pixel value, then YAFR_TEST (smooth) boils down to
* Catmull-Rom, and the computation of the correction is a waste.
* Extreme case: If all the pixels are either pure black or pure white
* in some region, as in some text images (more generally, if the
* region is "bichromatic"), then the YAFR_TEST (smooth) correction is 0 in
* the interior of the bichromatic region.
*/
/* Pointers to write to / read from, how much to add to move right a pixel,
* how much to add to move down a line.
*/
/* T is the type of pixels we are reading and writing.
* D is a type for calculation of the yafr correction: it needs to be large
* enough to hold squares of differences ... so for char types, int will work,
* for others we need float or even double.
*/
template <typename T, typename D> static inline void
catrom_yafr_test(
PEL *out, PEL *in,
const int channels,
const int pixels_per_buffer_row,
const float sharpening,
const float cardinal_one,
const float cardinal_two,
const float cardinal_thr,
const float cardinal_fou,
const float cardinal_uno,
const float cardinal_dos,
const float cardinal_tre,
const float cardinal_qua,
const float left_width_times_up__height_times_rite_width,
const float left_width_times_dow_height_times_rite_width,
const float left_width_times_up__height_times_dow_height,
const float rite_width_times_up__height_times_dow_height )
{
T* restrict out = (T *) pout;
const T* restrict in = (T *) pin;
/* "sharpening" is a continuous method parameter which is
* proportional to the amount of "diagonal straightening" which the
* nonlinear correction part of the method may add to the underlying
* linear scheme. You may also think of it as a sharpening
* parameter: higher values correspond to more sharpening, and
* negative values lead to strange looking effects.
*
* The default value is sharpening = 29/32 when the scheme being
* "straightened" is Catmull-Rom---as is the case here. This value
* fixes key pixel values near the diagonal boundary between two
* monochrome regions (the diagonal boundary pixel values being set
* to the halfway colour).
*
* If resampling seems to add unwanted texture artifacts, push
* sharpening toward 0. It is not generally not recommended to set
* sharpening to a value larger than 4.
*
* Sharpening is halved because the .5 which has to do with the
* relative coordinates of the evaluation points (which has to do
* with .5*rite_width etc) is folded into the constant to save
* flops. Consequently, the largest recommended value of
* sharpening_over_two is 2=4/2.
*
* In order to simplify interfacing with users, the parameter which
* should be set by the user is normalized so that user_sharpening =
* 1 when sharpening is equal to the recommended value. Consistently
* with the above discussion, values of user_sharpening between 0
* and about 3.625 give good results.
*/
const float sharpening_over_two = sharpening * 0.453125f;
/*
* The input pixel values are described by the following stencil.
* Spanish abbreviations are used to label positions from top to
* bottom, English ones to label positions from left to right,:
*
* (ix-1,iy-1) (ix,iy-1) (ix+1,iy-1) (ix+2,iy-1)
* =uno_one =uno_two =uno_thr = uno_fou
*
* (ix-1,iy) (ix,iy) (ix+1,iy) (ix+2,iy)
* =dos_one =dos_two =dos_thr = dos_fou
*
* (ix-1,iy+1) (ix,iy+1) (ix+1,iy+1) (ix+2,iy+1)
* =tre_one =tre_two =tre_thr = tre_fou
*
* (ix-1,iy+2) (ix,iy+2) (ix+1,iy+2) (ix+2,iy+2)
* =qua_one =qua_two =qua_thr = qua_fou
*/
/*
* Load the useful pixel values for the channel under
* consideration. The in pointer is assumed
* to point to uno_one when catrom_yafr_test is entered.
*/
const T uno_one = in[ 0 ];
const T uno_two = in[ channels];
const T uno_thr = in[ 2 * channels];
const T uno_fou = in[ 3 * channels];
const T dos_one = in[ pixels_per_buffer_row * channels];
const T dos_two = in[(1 + pixels_per_buffer_row) * channels];
const T dos_thr = in[(2 + pixels_per_buffer_row) * channels];
const T dos_fou = in[(3 + pixels_per_buffer_row) * channels];
const T tre_one = in[ 2 * pixels_per_buffer_row * channels];
const T tre_two = in[(1 + 2 * pixels_per_buffer_row) * channels];
const T tre_thr = in[(2 + 2 * pixels_per_buffer_row) * channels];
const T tre_fou = in[(3 + 2 * pixels_per_buffer_row) * channels];
const T qua_one = in[ 3 * pixels_per_buffer_row * channels];
const T qua_two = in[(1 + 3 * pixels_per_buffer_row) * channels];
const T qua_thr = in[(2 + 3 * pixels_per_buffer_row) * channels];
const T qua_fou = in[(3 + 3 * pixels_per_buffer_row) * channels];
/*
* Computation of the YAFR_TEST correction:
*
* Basically, if two consecutive pixel value differences have the
* same sign, the smallest one (in absolute value) is taken to be
* the corresponding slope. If they don't have the same sign, the
* corresponding slope is set to 0.
*
* Four such pairs (vertical and horizontal) of slopes need to be
* computed, one pair for each of the pixels which potentially
* overlap the unit area centered at the interpolation point.
*/
/*
* Beginning of the computation of the "up" horizontal slopes:
*/
const D prem__up = dos_two - dos_one;
const D deux__up = dos_thr - dos_two;
const D troi__up = dos_fou - dos_thr;
/*
* "down" horizontal slopes:
*/
const D prem_dow = tre_two - tre_one;
const D deux_dow = tre_thr - tre_two;
const D troi_dow = tre_fou - tre_thr;
/*
* "left" vertical slopes:
*/
const D prem_left = dos_two - uno_two;
const D deux_left = tre_two - dos_two;
const D troi_left = qua_two - tre_two;
/*
* "right" vertical slopes:
*/
const D prem_rite = dos_thr - uno_thr;
const D deux_rite = tre_thr - dos_thr;
const D troi_rite = qua_thr - tre_thr;
/*
* Back to "up":
*/
const D prem__up_squared = prem__up * prem__up;
const D deux__up_squared = deux__up * deux__up;
const D troi__up_squared = troi__up * troi__up;
/*
* Back to "down":
*/
const D prem_dow_squared = prem_dow * prem_dow;
const D deux_dow_squared = deux_dow * deux_dow;
const D troi_dow_squared = troi_dow * troi_dow;
/*
* Back to "left":
*/
const D prem_left_squared = prem_left * prem_left;
const D deux_left_squared = deux_left * deux_left;
const D troi_left_squared = troi_left * troi_left;
/*
* Back to "right":
*/
const D prem_rite_squared = prem_rite * prem_rite;
const D deux_rite_squared = deux_rite * deux_rite;
const D troi_rite_squared = troi_rite * troi_rite;
/*
* "up":
*/
const D prem__up_times_deux__up = prem__up * deux__up;
const D deux__up_times_troi__up = deux__up * troi__up;
/*
* "down":
*/
const D prem_dow_times_deux_dow = prem_dow * deux_dow;
const D deux_dow_times_troi_dow = deux_dow * troi_dow;
/*
* "left":
*/
const D prem_left_times_deux_left = prem_left * deux_left;
const D deux_left_times_troi_left = deux_left * troi_left;
/*
* "right":
*/
const D prem_rite_times_deux_rite = prem_rite * deux_rite;
const D deux_rite_times_troi_rite = deux_rite * troi_rite;
/*
* Branching parts of the computation of the YAFR_TEST correction (could
* be unbranched using arithmetic branching and C99 math intrinsics,
* although the compiler may be smart enough to remove the branching
* on its own):
*/
/*
* "up":
*/
const D prem__up_vs_deux__up =
prem__up_squared < deux__up_squared ? prem__up : deux__up;
const D deux__up_vs_troi__up =
deux__up_squared < troi__up_squared ? deux__up : troi__up;
/*
* "down":
*/
const D prem_dow_vs_deux_dow =
prem_dow_squared < deux_dow_squared ? prem_dow : deux_dow;
const D deux_dow_vs_troi_dow =
deux_dow_squared < troi_dow_squared ? deux_dow : troi_dow;
/*
* "left":
*/
const D prem_left_vs_deux_left =
prem_left_squared < deux_left_squared ? prem_left : deux_left;
const D deux_left_vs_troi_left =
deux_left_squared < troi_left_squared ? deux_left : troi_left;
/*
* "right":
*/
const D prem_rite_vs_deux_rite =
prem_rite_squared < deux_rite_squared ? prem_rite : deux_rite;
const D deux_rite_vs_troi_rite =
deux_rite_squared < troi_rite_squared ? deux_rite : troi_rite;
/*
* The YAFR_TEST correction computation will resume after the
* computation of the Catmull-Rom baseline.
*/
/*
* Catmull-Rom baseline contribution:
*/
const float catmull_rom =
cardinal_uno * (
cardinal_one * uno_one +
cardinal_two * uno_two +
cardinal_thr * uno_thr +
cardinal_fou * uno_fou
) +
cardinal_dos * (
cardinal_one * dos_one +
cardinal_two * dos_two +
cardinal_thr * dos_thr +
cardinal_fou * dos_fou
) +
cardinal_tre * (
cardinal_one * tre_one +
cardinal_two * tre_two +
cardinal_thr * tre_thr +
cardinal_fou * tre_fou
) +
cardinal_qua * (
cardinal_one * qua_one +
cardinal_two * qua_two +
cardinal_thr * qua_thr +
cardinal_fou * qua_fou
);
/*
* Computation of the YAFR_TEST slopes.
*/
/*
* "up":
*/
const D mx_left__up =
prem__up_times_deux__up < 0.f ? 0.f : prem__up_vs_deux__up;
const D mx_rite__up =
deux__up_times_troi__up < 0.f ? 0.f : deux__up_vs_troi__up;
/*
* "down":
*/
const D mx_left_dow =
prem_dow_times_deux_dow < 0.f ? 0.f : prem_dow_vs_deux_dow;
const D mx_rite_dow =
deux_dow_times_troi_dow < 0.f ? 0.f : deux_dow_vs_troi_dow;
/*
* "left":
*/
const D my_left__up =
prem_left_times_deux_left < 0.f ? 0.f : prem_left_vs_deux_left;
const D my_left_dow =
deux_left_times_troi_left < 0.f ? 0.f : deux_left_vs_troi_left;
/*
* "right":
*/
const D my_rite__up =
prem_rite_times_deux_rite < 0.f ? 0.f : prem_rite_vs_deux_rite;
const D my_rite_dow =
deux_rite_times_troi_rite < 0.f ? 0.f : deux_rite_vs_troi_rite;
/*
* Assemble the unweighted YAFR_TEST correction:
*/
const float unweighted_yafr_test_correction =
left_width_times_up__height_times_rite_width *
(mx_left__up - mx_rite__up) +
left_width_times_dow_height_times_rite_width *
(mx_left_dow - mx_rite_dow) +
left_width_times_up__height_times_dow_height *
(my_left__up - my_left_dow) +
rite_width_times_up__height_times_dow_height *
(my_rite__up - my_rite_dow);
/*
* Add the Catmull-Rom baseline and the weighted YAFR_TEST correction:
*/
const T newval =
sharpening_over_two * unweighted_yafr_test_correction +
catmull_rom;
*out = newval;
}
static void
vips_interpolate_yafr_test_interpolate( VipsInterpolate *interpolate,
REGION *out, REGION *in,
int out_x, int out_y, double x, double y )
{
VipsInterpolateYafrTest *yafr_test =
VIPS_INTERPOLATE_YAFR_TEST( interpolate );
/*
* Note: The computation is structured to foster software
* pipelining.
*/
/*
* x is understood to increase from left to right, y, from top to
* bottom. Consequently, ix and iy are the indices of the pixel
* located at or to the left, and at or above. the sampling point.
*
* floor is used to make sure that the transition through 0 is
* smooth. If it is known that negative x and y will never be used,
* cast (which truncates) could be used instead.
*/
const gint ix = FLOOR (x);
const gint iy = FLOOR (y);
/*
* Each (channel's) output pixel value is obtained by combining four
* "pieces," each piece corresponding to the set of points which are
* closest to the four pixels closest to the (x,y) position, pixel
* positions which have coordinates and labels as follows:
*
* (ix,iy) (ix+1,iy)
* =left__up =rite__up
*
* <- (x,y) is somewhere in the convex hull
*
* (ix,iy+1) (ix+1,iy+1)
* =left_dow =rite_dow
*/
/*
* rite_width is the width of the overlaps of the unit averaging box
* (which is centered at the position where an interpolated value is
* desired), with the closest unit pixel areas to the right.
*
* left_width is the width of the overlaps of the unit averaging box
* (which is centered at the position where an interpolated value is
* desired), with the closest unit pixel areas to the left.
*/
const float rite_width = x - ix;
const float dow_height = y - iy;
const float left_width = 1.f - rite_width;
const float up__height = 1.f - dow_height;
/*
* .5*rite_width is the x-coordinate of the center of the overlap of
* the averaging box with the left pixel areas, relative to the
* position of the centers of the left pixels.
*
* -.5*left_width is the x-coordinate ... right pixel areas,
* relative to ... the right pixels.
*
* .5*dow_height is the y-coordinate of the center of the overlap
* of the averaging box with the up pixel areas, relative to the
* position of the centers of the up pixels.
*
* -.5*up__height is the y-coordinate ... down pixel areas, relative
* to ... the down pixels.
*/
const float left_width_times_rite_width = left_width * rite_width;
const float up__height_times_dow_height = up__height * dow_height;
const float cardinal_two =
left_width_times_rite_width * (-1.5f * rite_width + 1.f) +
left_width;
const float cardinal_dos =
up__height_times_dow_height * (-1.5f * dow_height + 1.f) +
up__height;
const float minus_half_left_width_times_rite_width =
-.5f * left_width_times_rite_width;
const float minus_half_up__height_times_dow_height =
-.5f * up__height_times_dow_height;
const float left_width_times_up__height_times_rite_width =
left_width_times_rite_width * up__height;
const float left_width_times_dow_height_times_rite_width =
left_width_times_rite_width * dow_height;
const float left_width_times_up__height_times_dow_height =
up__height_times_dow_height * left_width;
const float rite_width_times_up__height_times_dow_height =
up__height_times_dow_height * rite_width;
const float cardinal_one =
minus_half_left_width_times_rite_width * left_width;
const float cardinal_uno =
minus_half_up__height_times_dow_height * up__height;
const float cardinal_fou =
minus_half_left_width_times_rite_width * rite_width;
const float cardinal_qua =
minus_half_up__height_times_dow_height * dow_height;
const float cardinal_thr =
1.f - (minus_half_left_width_times_rite_width + cardinal_two);
const float cardinal_tre =
1.f - (minus_half_up__height_times_dow_height + cardinal_dos);
/*
* Set the tile pointer to the first relevant value. Since the
* pointer initially points to dos_two, we need to rewind it one
* tile row, then go back one additional pixel.
*/
const PEL *p = (PEL *) IM_REGION_ADDR( in, ix - 1, iy - 1 );
/* Pel size and line size.
*/
const int channels = in->im->Bands;
const int pixels_per_buffer_row =
IM_REGION_LSKIP( in ) / (sizeof( float ) * channels);
/* Where we write the result.
*/
PEL *q = (PEL *) IM_REGION_ADDR( out, out_x, out_y );
/* Put this in a macro to save some typing.
*/
#define CALL(T, D) \
catrom_yafr_test<T, D>(q + z, p + z, \
channels, pixels_per_buffer_row, \
yafr_test->sharpening, \
cardinal_one, \
cardinal_two, \
cardinal_thr, \
cardinal_fou, \
cardinal_uno, \
cardinal_dos, \
cardinal_tre, \
cardinal_qua, \
left_width_times_up__height_times_rite_width, \
left_width_times_dow_height_times_rite_width, \
left_width_times_up__height_times_dow_height, \
rite_width_times_up__height_times_dow_height);
switch( in->im->BandFmt ) {
case IM_BANDFMT_UCHAR:
for( int z = 0; z < channels; z++ )
CALL( unsigned char, int );
break;
case IM_BANDFMT_CHAR:
for( int z = 0; z < channels; z++ )
CALL( signed char, int );
break;
case IM_BANDFMT_USHORT:
for( int z = 0; z < channels; z++ )
CALL( unsigned short, float );
break;
case IM_BANDFMT_SHORT:
for( int z = 0; z < channels; z++ )
CALL( signed short, float );
break;
case IM_BANDFMT_UINT:
for( int z = 0; z < channels; z++ )
CALL( unsigned int, float );
break;
case IM_BANDFMT_INT:
for( int z = 0; z < channels; z++ )
CALL( signed int, float );
break;
case IM_BANDFMT_FLOAT:
for( int z = 0; z < channels; z++ )
CALL( float, float );
break;
case IM_BANDFMT_DOUBLE:
for( int z = 0; z < channels; z++ )
CALL( float, float );
break;
default:
}
}
static void
vips_interpolate_yafr_test_class_init( VipsInterpolateYafrTestClass *class )
{
VipsInterpolateClass *interpolate_class =
VIPS_INTERPOLATE_CLASS( class );
vips_interpolate_yafr_test_parent_class =
g_type_class_peek_parent( class );
interpolate_class->interpolate = vips_interpolate_yafr_test_interpolate;
interpolate_class->window_size = 4;
}
static void
vips_interpolate_yafr_test_init( VipsInterpolateYafrTest *yafr_test )
{
#ifdef DEBUG
printf( "vips_interpolate_yafr_test_init: " );
vips_object_print( VIPS_OBJECT( yafr_test ) );
#endif /*DEBUG*/
yafr_test->sharpening = 1.0;
}
GType
vips_interpolate_yafr_test_get_type( void )
{
static GType type = 0;
if( !type ) {
static const GTypeInfo info = {
sizeof( VipsInterpolateYafrTestClass ),
NULL, /* base_init */
NULL, /* base_finalize */
(GClassInitFunc) vips_interpolate_yafr_test_class_init,
NULL, /* class_finalize */
NULL, /* class_data */
sizeof( VipsInterpolateYafrTest ),
32, /* n_preallocs */
(GInstanceInitFunc) vips_interpolate_yafr_test_init,
};
type = g_type_register_static( VIPS_TYPE_INTERPOLATE,
"VipsInterpolateYafrTest", &info, 0 );
}
return( type );
}
VipsInterpolate *
vips_interpolate_yafr_test_new( void )
{
return( VIPS_INTERPOLATE( g_object_new(
VIPS_TYPE_INTERPOLATE_YAFR_TEST, NULL ) ) );
}
void
vips_interpolate_yafr_test_set_sharpening( VipsInterpolateYafrTest *yafr_test,
double sharpening )
{
yafr_test->sharpening = sharpening;
}
/* Convenience: return a static yafr_test you don't need to free.
*/
VipsInterpolate *
vips_interpolate_yafr_test_static( void )
{
static VipsInterpolate *interpolate = NULL;
if( !interpolate )
interpolate = vips_interpolate_yafr_test_new();
return( interpolate );
}