1 // xImaInt.cpp : interpolation functions
\r
2 /* 02/2004 - Branko Brevensek
\r
3 * CxImage version 6.0.0 02/Feb/2008 - Davide Pizzolato - www.xdp.it
\r
9 #if CXIMAGE_SUPPORT_INTERPOLATION
\r
11 ////////////////////////////////////////////////////////////////////////////////
\r
13 * Recalculates coordinates according to specified overflow method.
\r
14 * If pixel (x,y) lies within image, nothing changes.
\r
16 * \param x, y - coordinates of pixel
\r
17 * \param ofMethod - overflow method
\r
19 * \return x, y - new coordinates (pixel (x,y) now lies inside image)
\r
21 * \author ***bd*** 2.2004
\r
23 void CxImage::OverflowCoordinates(long &x, long &y, OverflowMethod const ofMethod)
\r
25 if (IsInside(x,y)) return; //if pixel is within bounds, no change
\r
29 x=max(x,0); x=min(x, head.biWidth-1);
\r
30 y=max(y,0); y=min(y, head.biHeight-1);
\r
34 x = x % head.biWidth;
\r
35 y = y % head.biHeight;
\r
36 if (x<0) x = head.biWidth + x;
\r
37 if (y<0) y = head.biHeight + y;
\r
40 //mirror pixels near border
\r
41 if (x<0) x=((-x) % head.biWidth);
\r
42 else if (x>=head.biWidth) x=head.biWidth-(x % head.biWidth + 1);
\r
43 if (y<0) y=((-y) % head.biHeight);
\r
44 else if (y>=head.biHeight) y=head.biHeight-(y % head.biHeight + 1);
\r
51 ////////////////////////////////////////////////////////////////////////////////
\r
53 * See OverflowCoordinates for integer version
\r
54 * \author ***bd*** 2.2004
\r
56 void CxImage::OverflowCoordinates(float &x, float &y, OverflowMethod const ofMethod)
\r
58 if (x>=0 && x<head.biWidth && y>=0 && y<head.biHeight) return; //if pixel is within bounds, no change
\r
62 x=max(x,0); x=min(x, head.biWidth-1);
\r
63 y=max(y,0); y=min(y, head.biHeight-1);
\r
67 x = (float)fmod(x, (float) head.biWidth);
\r
68 y = (float)fmod(y, (float) head.biHeight);
\r
69 if (x<0) x = head.biWidth + x;
\r
70 if (y<0) y = head.biHeight + y;
\r
73 //mirror pixels near border
\r
74 if (x<0) x=(float)fmod(-x, (float) head.biWidth);
\r
75 else if (x>=head.biWidth) x=head.biWidth-((float)fmod(x, (float) head.biWidth) + 1);
\r
76 if (y<0) y=(float)fmod(-y, (float) head.biHeight);
\r
77 else if (y>=head.biHeight) y=head.biHeight-((float)fmod(y, (float) head.biHeight) + 1);
\r
84 ////////////////////////////////////////////////////////////////////////////////
\r
86 * Method return pixel color. Different methods are implemented for out of bounds pixels.
\r
87 * If an image has alpha channel, alpha value is returned in .RGBReserved.
\r
89 * \param x,y : pixel coordinates
\r
90 * \param ofMethod : out-of-bounds method:
\r
91 * - OF_WRAP - wrap over to pixels on other side of the image
\r
92 * - OF_REPEAT - repeat last pixel on the edge
\r
93 * - OF_COLOR - return input value of color
\r
94 * - OF_BACKGROUND - return background color (if not set, return input color)
\r
95 * - OF_TRANSPARENT - return transparent pixel
\r
97 * \param rplColor : input color (returned for out-of-bound coordinates in OF_COLOR mode and if other mode is not applicable)
\r
99 * \return color : color of pixel
\r
100 * \author ***bd*** 2.2004
\r
102 RGBQUAD CxImage::GetPixelColorWithOverflow(long x, long y, OverflowMethod const ofMethod, RGBQUAD* const rplColor)
\r
104 RGBQUAD color; //color to return
\r
105 if ((!IsInside(x,y)) || pDib==NULL) { //is pixel within bouns?:
\r
106 //pixel is out of bounds or no DIB
\r
107 if (rplColor!=NULL)
\r
110 color.rgbRed=color.rgbGreen=color.rgbBlue=255; color.rgbReserved=0; //default replacement colour: white transparent
\r
112 if (pDib==NULL) return color;
\r
113 //pixel is out of bounds:
\r
114 switch (ofMethod) {
\r
115 case OM_TRANSPARENT:
\r
116 #if CXIMAGE_SUPPORT_ALPHA
\r
117 if (AlphaIsValid()) {
\r
118 //alpha transparency is supported and image has alpha layer
\r
119 color.rgbReserved=0;
\r
121 #endif //CXIMAGE_SUPPORT_ALPHA
\r
122 //no alpha transparency
\r
123 if (GetTransIndex()>=0) {
\r
124 color=GetTransColor(); //single color transparency enabled (return transparent color)
\r
126 #if CXIMAGE_SUPPORT_ALPHA
\r
128 #endif //CXIMAGE_SUPPORT_ALPHA
\r
130 case OM_BACKGROUND:
\r
131 //return background color (if it exists, otherwise input value)
\r
132 if (info.nBkgndIndex >= 0) {
\r
133 if (head.biBitCount<24) color = GetPaletteColor((BYTE)info.nBkgndIndex);
\r
134 else color = info.nBkgndColor;
\r
140 OverflowCoordinates(x,y,ofMethod);
\r
143 //simply return replacement color (OM_COLOR and others)
\r
147 //just return specified pixel (it's within bounds)
\r
148 return BlindGetPixelColor(x,y);
\r
151 ////////////////////////////////////////////////////////////////////////////////
\r
153 * This method reconstructs image according to chosen interpolation method and then returns pixel (x,y).
\r
154 * (x,y) can lie between actual image pixels. If (x,y) lies outside of image, method returns value
\r
155 * according to overflow method.
\r
156 * This method is very useful for geometrical image transformations, where destination pixel
\r
157 * can often assume color value lying between source pixels.
\r
159 * \param (x,y) - coordinates of pixel to return
\r
160 * GPCI method recreates "analogue" image back from digital data, so x and y
\r
161 * are float values and color value of point (1.1,1) will generally not be same
\r
162 * as (1,1). Center of first pixel is at (0,0) and center of pixel right to it is (1,0).
\r
163 * (0.5,0) is half way between these two pixels.
\r
164 * \param inMethod - interpolation (reconstruction) method (kernel) to use:
\r
165 * - IM_NEAREST_NEIGHBOUR - returns colour of nearest lying pixel (causes stairy look of
\r
166 * processed images)
\r
167 * - IM_BILINEAR - interpolates colour from four neighbouring pixels (softens image a bit)
\r
168 * - IM_BICUBIC - interpolates from 16 neighbouring pixels (can produce "halo" artifacts)
\r
169 * - IM_BICUBIC2 - interpolates from 16 neighbouring pixels (perhaps a bit less halo artifacts
\r
171 * - IM_BSPLINE - interpolates from 16 neighbouring pixels (softens image, washes colours)
\r
172 * (As far as I know, image should be prefiltered for this method to give
\r
173 * good results... some other time :) )
\r
174 * This method uses bicubic interpolation kernel from CXImage 5.99a and older
\r
176 * - IM_LANCZOS - interpolates from 12*12 pixels (slow, ringing artifacts)
\r
178 * \param ofMethod - overflow method (see comments at GetPixelColorWithOverflow)
\r
179 * \param rplColor - pointer to color used for out of borders pixels in OM_COLOR mode
\r
180 * (and other modes if colour can't calculated in a specified way)
\r
182 * \return interpolated color value (including interpolated alpha value, if image has alpha layer)
\r
184 * \author ***bd*** 2.2004
\r
186 RGBQUAD CxImage::GetPixelColorInterpolated(
\r
188 InterpolationMethod const inMethod,
\r
189 OverflowMethod const ofMethod,
\r
190 RGBQUAD* const rplColor)
\r
192 //calculate nearest pixel
\r
193 int xi=(int)(x); if (x<0) xi--; //these replace (incredibly slow) floor (Visual c++ 2003, AMD Athlon)
\r
194 int yi=(int)(y); if (y<0) yi--;
\r
195 RGBQUAD color; //calculated colour
\r
197 switch (inMethod) {
\r
198 case IM_NEAREST_NEIGHBOUR:
\r
199 return GetPixelColorWithOverflow((long)(x+0.5f), (long)(y+0.5f), ofMethod, rplColor);
\r
201 //IM_BILINEAR: bilinear interpolation
\r
202 if (xi<-1 || xi>=head.biWidth || yi<-1 || yi>=head.biHeight) { //all 4 points are outside bounds?:
\r
203 switch (ofMethod) {
\r
204 case OM_COLOR: case OM_TRANSPARENT: case OM_BACKGROUND:
\r
205 //we don't need to interpolate anything with all points outside in this case
\r
206 return GetPixelColorWithOverflow(-999, -999, ofMethod, rplColor);
\r
208 //recalculate coordinates and use faster method later on
\r
209 OverflowCoordinates(x,y,ofMethod);
\r
210 xi=(int)(x); if (x<0) xi--; //x and/or y have changed ... recalculate xi and yi
\r
211 yi=(int)(y); if (y<0) yi--;
\r
214 //get four neighbouring pixels
\r
215 if ((xi+1)<head.biWidth && xi>=0 && (yi+1)<head.biHeight && yi>=0 && head.biClrUsed==0) {
\r
216 //all pixels are inside RGB24 image... optimize reading (and use fixed point arithmetic)
\r
217 WORD wt1=(WORD)((x-xi)*256.0f), wt2=(WORD)((y-yi)*256.0f);
\r
218 WORD wd=wt1*wt2>>8;
\r
221 WORD wa=256-wt1-wc;
\r
223 BYTE *pxptr=(BYTE*)info.pImage+yi*info.dwEffWidth+xi*3;
\r
224 wbb=wa*(*pxptr++); wgg=wa*(*pxptr++); wrr=wa*(*pxptr++);
\r
225 wbb+=wb*(*pxptr++); wgg+=wb*(*pxptr++); wrr+=wb*(*pxptr);
\r
226 pxptr+=(info.dwEffWidth-5); //move to next row
\r
227 wbb+=wc*(*pxptr++); wgg+=wc*(*pxptr++); wrr+=wc*(*pxptr++);
\r
228 wbb+=wd*(*pxptr++); wgg+=wd*(*pxptr++); wrr+=wd*(*pxptr);
\r
229 color.rgbRed=(BYTE) (wrr>>8); color.rgbGreen=(BYTE) (wgg>>8); color.rgbBlue=(BYTE) (wbb>>8);
\r
230 #if CXIMAGE_SUPPORT_ALPHA
\r
233 //image has alpha layer... we have to do the same for alpha data
\r
234 pxptr=AlphaGetPointer(xi,yi); //pointer to first byte
\r
235 waa=wa*(*pxptr++); waa+=wb*(*pxptr); //first two pixels
\r
236 pxptr+=(head.biWidth-1); //move to next row
\r
237 waa+=wc*(*pxptr++); waa+=wd*(*pxptr); //and second row pixels
\r
238 color.rgbReserved=(BYTE) (waa>>8);
\r
241 { //Alpha not supported or no alpha at all
\r
242 color.rgbReserved = 0;
\r
246 //default (slower) way to get pixels (not RGB24 or some pixels out of borders)
\r
247 float t1=x-xi, t2=y-yi;
\r
252 RGBQUAD rgb11,rgb21,rgb12,rgb22;
\r
253 rgb11=GetPixelColorWithOverflow(xi, yi, ofMethod, rplColor);
\r
254 rgb21=GetPixelColorWithOverflow(xi+1, yi, ofMethod, rplColor);
\r
255 rgb12=GetPixelColorWithOverflow(xi, yi+1, ofMethod, rplColor);
\r
256 rgb22=GetPixelColorWithOverflow(xi+1, yi+1, ofMethod, rplColor);
\r
257 //calculate linear interpolation
\r
258 color.rgbRed=(BYTE) (a*rgb11.rgbRed+b*rgb21.rgbRed+c*rgb12.rgbRed+d*rgb22.rgbRed);
\r
259 color.rgbGreen=(BYTE) (a*rgb11.rgbGreen+b*rgb21.rgbGreen+c*rgb12.rgbGreen+d*rgb22.rgbGreen);
\r
260 color.rgbBlue=(BYTE) (a*rgb11.rgbBlue+b*rgb21.rgbBlue+c*rgb12.rgbBlue+d*rgb22.rgbBlue);
\r
261 #if CXIMAGE_SUPPORT_ALPHA
\r
262 if (AlphaIsValid())
\r
263 color.rgbReserved=(BYTE) (a*rgb11.rgbReserved+b*rgb21.rgbReserved+c*rgb12.rgbReserved+d*rgb22.rgbReserved);
\r
266 { //Alpha not supported or no alpha at all
\r
267 color.rgbReserved = 0;
\r
287 //bicubic interpolation(s)
\r
288 if (((xi+2)<0) || ((xi-1)>=head.biWidth) || ((yi+2)<0) || ((yi-1)>=head.biHeight)) { //all points are outside bounds?:
\r
289 switch (ofMethod) {
\r
290 case OM_COLOR: case OM_TRANSPARENT: case OM_BACKGROUND:
\r
291 //we don't need to interpolate anything with all points outside in this case
\r
292 return GetPixelColorWithOverflow(-999, -999, ofMethod, rplColor);
\r
295 //recalculate coordinates and use faster method later on
\r
296 OverflowCoordinates(x,y,ofMethod);
\r
297 xi=(int)(x); if (x<0) xi--; //x and/or y have changed ... recalculate xi and yi
\r
298 yi=(int)(y); if (y<0) yi--;
\r
302 //some variables needed from here on
\r
303 int xii,yii; //x any y integer indexes for loops
\r
304 float kernel, kernelyc; //kernel cache
\r
305 float kernelx[12], kernely[4]; //precalculated kernel values
\r
306 float rr,gg,bb,aa; //accumulated color values
\r
307 //calculate multiplication factors for all pixels
\r
309 switch (inMethod) {
\r
311 for (i=0; i<4; i++) {
\r
312 kernelx[i]=KernelCubic((float)(xi+i-1-x));
\r
313 kernely[i]=KernelCubic((float)(yi+i-1-y));
\r
317 for (i=0; i<4; i++) {
\r
318 kernelx[i]=KernelGeneralizedCubic((float)(xi+i-1-x), -0.5);
\r
319 kernely[i]=KernelGeneralizedCubic((float)(yi+i-1-y), -0.5);
\r
323 for (i=0; i<4; i++) {
\r
324 kernelx[i]=KernelBSpline((float)(xi+i-1-x));
\r
325 kernely[i]=KernelBSpline((float)(yi+i-1-y));
\r
329 for (i=0; i<4; i++) {
\r
330 kernelx[i]=KernelBox((float)(xi+i-1-x));
\r
331 kernely[i]=KernelBox((float)(yi+i-1-y));
\r
335 for (i=0; i<4; i++) {
\r
336 kernelx[i]=KernelHermite((float)(xi+i-1-x));
\r
337 kernely[i]=KernelHermite((float)(yi+i-1-y));
\r
341 for (i=0; i<4; i++) {
\r
342 kernelx[i]=KernelHamming((float)(xi+i-1-x));
\r
343 kernely[i]=KernelHamming((float)(yi+i-1-y));
\r
347 for (i=0; i<4; i++) {
\r
348 kernelx[i]=KernelSinc((float)(xi+i-1-x));
\r
349 kernely[i]=KernelSinc((float)(yi+i-1-y));
\r
353 for (i=0; i<4; i++) {
\r
354 kernelx[i]=KernelBlackman((float)(xi+i-1-x));
\r
355 kernely[i]=KernelBlackman((float)(yi+i-1-y));
\r
359 for (i=0; i<4; i++) {
\r
360 kernelx[i]=KernelBessel((float)(xi+i-1-x));
\r
361 kernely[i]=KernelBessel((float)(yi+i-1-y));
\r
365 for (i=0; i<4; i++) {
\r
366 kernelx[i]=KernelGaussian((float)(xi+i-1-x));
\r
367 kernely[i]=KernelGaussian((float)(yi+i-1-y));
\r
371 for (i=0; i<4; i++) {
\r
372 kernelx[i]=KernelQuadratic((float)(xi+i-1-x));
\r
373 kernely[i]=KernelQuadratic((float)(yi+i-1-y));
\r
377 for (i=0; i<4; i++) {
\r
378 kernelx[i]=KernelMitchell((float)(xi+i-1-x));
\r
379 kernely[i]=KernelMitchell((float)(yi+i-1-y));
\r
383 for (i=0; i<4; i++) {
\r
384 kernelx[i]=KernelCatrom((float)(xi+i-1-x));
\r
385 kernely[i]=KernelCatrom((float)(yi+i-1-y));
\r
389 for (i=0; i<4; i++) {
\r
390 kernelx[i]=KernelHanning((float)(xi+i-1-x));
\r
391 kernely[i]=KernelHanning((float)(yi+i-1-y));
\r
395 for (i=0; i<4; i++) {
\r
396 kernelx[i]=KernelPower((float)(xi+i-1-x));
\r
397 kernely[i]=KernelPower((float)(yi+i-1-y));
\r
402 if (((xi+2)<head.biWidth) && xi>=1 && ((yi+2)<head.biHeight) && (yi>=1) && !IsIndexed()) {
\r
403 //optimized interpolation (faster pixel reads) for RGB24 images with all pixels inside bounds
\r
404 BYTE *pxptr, *pxptra;
\r
405 for (yii=yi-1; yii<yi+3; yii++) {
\r
406 pxptr=(BYTE *)BlindGetPixelPointer(xi-1, yii); //calculate pointer to first byte in row
\r
407 kernelyc=kernely[yii-(yi-1)];
\r
408 #if CXIMAGE_SUPPORT_ALPHA
\r
409 if (AlphaIsValid()) {
\r
410 //alpha is supported and valid (optimized bicubic int. for image with alpha)
\r
411 pxptra=AlphaGetPointer(xi-1, yii);
\r
412 kernel=kernelyc*kernelx[0];
\r
413 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++); aa+=kernel*(*pxptra++);
\r
414 kernel=kernelyc*kernelx[1];
\r
415 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++); aa+=kernel*(*pxptra++);
\r
416 kernel=kernelyc*kernelx[2];
\r
417 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++); aa+=kernel*(*pxptra++);
\r
418 kernel=kernelyc*kernelx[3];
\r
419 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr); aa+=kernel*(*pxptra);
\r
422 //alpha not supported or valid (optimized bicubic int. for no alpha channel)
\r
424 kernel=kernelyc*kernelx[0];
\r
425 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++);
\r
426 kernel=kernelyc*kernelx[1];
\r
427 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++);
\r
428 kernel=kernelyc*kernelx[2];
\r
429 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++);
\r
430 kernel=kernelyc*kernelx[3];
\r
431 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr);
\r
435 //slower more flexible interpolation for border pixels and paletted images
\r
437 for (yii=yi-1; yii<yi+3; yii++) {
\r
438 kernelyc=kernely[yii-(yi-1)];
\r
439 for (xii=xi-1; xii<xi+3; xii++) {
\r
440 kernel=kernelyc*kernelx[xii-(xi-1)];
\r
441 rgbs=GetPixelColorWithOverflow(xii, yii, ofMethod, rplColor);
\r
442 rr+=kernel*rgbs.rgbRed;
\r
443 gg+=kernel*rgbs.rgbGreen;
\r
444 bb+=kernel*rgbs.rgbBlue;
\r
445 #if CXIMAGE_SUPPORT_ALPHA
\r
446 aa+=kernel*rgbs.rgbReserved;
\r
451 //for all colors, clip to 0..255 and assign to RGBQUAD
\r
452 if (rr>255) rr=255; if (rr<0) rr=0; color.rgbRed=(BYTE) rr;
\r
453 if (gg>255) gg=255; if (gg<0) gg=0; color.rgbGreen=(BYTE) gg;
\r
454 if (bb>255) bb=255; if (bb<0) bb=0; color.rgbBlue=(BYTE) bb;
\r
455 #if CXIMAGE_SUPPORT_ALPHA
\r
456 if (AlphaIsValid()) {
\r
457 if (aa>255) aa=255; if (aa<0) aa=0; color.rgbReserved=(BYTE) aa;
\r
460 { //Alpha not supported or no alpha at all
\r
461 color.rgbReserved = 0;
\r
465 //lanczos window (16*16) sinc interpolation
\r
466 if (((xi+6)<0) || ((xi-5)>=head.biWidth) || ((yi+6)<0) || ((yi-5)>=head.biHeight)) {
\r
467 //all points are outside bounds
\r
468 switch (ofMethod) {
\r
469 case OM_COLOR: case OM_TRANSPARENT: case OM_BACKGROUND:
\r
470 //we don't need to interpolate anything with all points outside in this case
\r
471 return GetPixelColorWithOverflow(-999, -999, ofMethod, rplColor);
\r
474 //recalculate coordinates and use faster method later on
\r
475 OverflowCoordinates(x,y,ofMethod);
\r
476 xi=(int)(x); if (x<0) xi--; //x and/or y have changed ... recalculate xi and yi
\r
477 yi=(int)(y); if (y<0) yi--;
\r
481 for (xii=xi-5; xii<xi+7; xii++) kernelx[xii-(xi-5)]=KernelLanczosSinc((float)(xii-x), 6.0f);
\r
484 if (((xi+6)<head.biWidth) && ((xi-5)>=0) && ((yi+6)<head.biHeight) && ((yi-5)>=0) && !IsIndexed()) {
\r
485 //optimized interpolation (faster pixel reads) for RGB24 images with all pixels inside bounds
\r
486 BYTE *pxptr, *pxptra;
\r
487 for (yii=yi-5; yii<yi+7; yii++) {
\r
488 pxptr=(BYTE *)BlindGetPixelPointer(xi-5, yii); //calculate pointer to first byte in row
\r
489 kernelyc=KernelLanczosSinc((float)(yii-y),6.0f);
\r
490 #if CXIMAGE_SUPPORT_ALPHA
\r
491 if (AlphaIsValid()) {
\r
492 //alpha is supported and valid
\r
493 pxptra=AlphaGetPointer(xi-1, yii);
\r
494 for (xii=0; xii<12; xii++) {
\r
495 kernel=kernelyc*kernelx[xii];
\r
496 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++); aa+=kernel*(*pxptra++);
\r
500 //alpha not supported or valid
\r
502 for (xii=0; xii<12; xii++) {
\r
503 kernel=kernelyc*kernelx[xii];
\r
504 bb+=kernel*(*pxptr++); gg+=kernel*(*pxptr++); rr+=kernel*(*pxptr++);
\r
509 //slower more flexible interpolation for border pixels and paletted images
\r
511 for (yii=yi-5; yii<yi+7; yii++) {
\r
512 kernelyc=KernelLanczosSinc((float)(yii-y),6.0f);
\r
513 for (xii=xi-5; xii<xi+7; xii++) {
\r
514 kernel=kernelyc*kernelx[xii-(xi-5)];
\r
515 rgbs=GetPixelColorWithOverflow(xii, yii, ofMethod, rplColor);
\r
516 rr+=kernel*rgbs.rgbRed;
\r
517 gg+=kernel*rgbs.rgbGreen;
\r
518 bb+=kernel*rgbs.rgbBlue;
\r
519 #if CXIMAGE_SUPPORT_ALPHA
\r
520 aa+=kernel*rgbs.rgbReserved;
\r
525 //for all colors, clip to 0..255 and assign to RGBQUAD
\r
526 if (rr>255) rr=255; if (rr<0) rr=0; color.rgbRed=(BYTE) rr;
\r
527 if (gg>255) gg=255; if (gg<0) gg=0; color.rgbGreen=(BYTE) gg;
\r
528 if (bb>255) bb=255; if (bb<0) bb=0; color.rgbBlue=(BYTE) bb;
\r
529 #if CXIMAGE_SUPPORT_ALPHA
\r
530 if (AlphaIsValid()) {
\r
531 if (aa>255) aa=255; if (aa<0) aa=0; color.rgbReserved=(BYTE) aa;
\r
534 { //Alpha not supported or no alpha at all
\r
535 color.rgbReserved = 0;
\r
540 ////////////////////////////////////////////////////////////////////////////////
\r
542 * Helper function for GetAreaColorInterpolated.
\r
543 * Adds 'surf' portion of image pixel with color 'color' to (rr,gg,bb,aa).
\r
545 void CxImage::AddAveragingCont(RGBQUAD const &color, float const surf, float &rr, float &gg, float &bb, float &aa)
\r
547 rr+=color.rgbRed*surf;
\r
548 gg+=color.rgbGreen*surf;
\r
549 bb+=color.rgbBlue*surf;
\r
550 #if CXIMAGE_SUPPORT_ALPHA
\r
551 aa+=color.rgbReserved*surf;
\r
554 ////////////////////////////////////////////////////////////////////////////////
\r
556 * This method is similar to GetPixelColorInterpolated, but this method also properly handles
\r
558 * If you need to sample original image with interval of more than 1 pixel (as when shrinking an image),
\r
559 * you should use this method instead of GetPixelColorInterpolated or aliasing will occur.
\r
560 * When area width and height are both less than pixel, this method gets pixel color by interpolating
\r
561 * color of frame center with selected (inMethod) interpolation by calling GetPixelColorInterpolated.
\r
562 * If width and height are more than 1, method calculates color by averaging color of pixels within area.
\r
563 * Interpolation method is not used in this case. Pixel color is interpolated by averaging instead.
\r
564 * If only one of both is more than 1, method uses combination of interpolation and averaging.
\r
565 * Chosen interpolation method is used, but since it is averaged later on, there is little difference
\r
566 * between IM_BILINEAR (perhaps best for this case) and better methods. IM_NEAREST_NEIGHBOUR again
\r
567 * leads to aliasing artifacts.
\r
568 * This method is a bit slower than GetPixelColorInterpolated and when aliasing is not a problem, you should
\r
569 * simply use the later.
\r
571 * \param xc, yc - center of (rectangular) area
\r
572 * \param w, h - width and height of area
\r
573 * \param inMethod - interpolation method that is used, when interpolation is used (see above)
\r
574 * \param ofMethod - overflow method used when retrieving individual pixel colors
\r
575 * \param rplColor - replacement colour to use, in OM_COLOR
\r
577 * \author ***bd*** 2.2004
\r
579 RGBQUAD CxImage::GetAreaColorInterpolated(
\r
580 float const xc, float const yc, float const w, float const h,
\r
581 InterpolationMethod const inMethod,
\r
582 OverflowMethod const ofMethod,
\r
583 RGBQUAD* const rplColor)
\r
585 RGBQUAD color; //calculated colour
\r
587 if (h<=1 && w<=1) {
\r
588 //both width and height are less than one... we will use interpolation of center point
\r
589 return GetPixelColorInterpolated(xc, yc, inMethod, ofMethod, rplColor);
\r
591 //area is wider and/or taller than one pixel:
\r
592 CxRect2 area(xc-w/2.0f, yc-h/2.0f, xc+w/2.0f, yc+h/2.0f); //area
\r
593 int xi1=(int)(area.botLeft.x+0.49999999f); //low x
\r
594 int yi1=(int)(area.botLeft.y+0.49999999f); //low y
\r
597 int xi2=(int)(area.topRight.x+0.5f); //top x
\r
598 int yi2=(int)(area.topRight.y+0.5f); //top y (for loops)
\r
600 float rr,gg,bb,aa; //red, green, blue and alpha components
\r
602 int x,y; //loop counters
\r
603 float s=0; //surface of all pixels
\r
604 float cps; //surface of current crosssection
\r
606 //width and height of area are greater than one pixel, so we can employ "ordinary" averaging
\r
607 CxRect2 intBL, intTR; //bottom left and top right intersection
\r
608 intBL=area.CrossSection(CxRect2(((float)xi1)-0.5f, ((float)yi1)-0.5f, ((float)xi1)+0.5f, ((float)yi1)+0.5f));
\r
609 intTR=area.CrossSection(CxRect2(((float)xi2)-0.5f, ((float)yi2)-0.5f, ((float)xi2)+0.5f, ((float)yi2)+0.5f));
\r
610 float wBL, wTR, hBL, hTR;
\r
611 wBL=intBL.Width(); //width of bottom left pixel-area intersection
\r
612 hBL=intBL.Height(); //height of bottom left...
\r
613 wTR=intTR.Width(); //width of top right...
\r
614 hTR=intTR.Height(); //height of top right...
\r
616 AddAveragingCont(GetPixelColorWithOverflow(xi1,yi1,ofMethod,rplColor), wBL*hBL, rr, gg, bb, aa); //bottom left pixel
\r
617 AddAveragingCont(GetPixelColorWithOverflow(xi2,yi1,ofMethod,rplColor), wTR*hBL, rr, gg, bb, aa); //bottom right pixel
\r
618 AddAveragingCont(GetPixelColorWithOverflow(xi1,yi2,ofMethod,rplColor), wBL*hTR, rr, gg, bb, aa); //top left pixel
\r
619 AddAveragingCont(GetPixelColorWithOverflow(xi2,yi2,ofMethod,rplColor), wTR*hTR, rr, gg, bb, aa); //top right pixel
\r
620 //bottom and top row
\r
621 for (x=xi1+1; x<xi2; x++) {
\r
622 AddAveragingCont(GetPixelColorWithOverflow(x,yi1,ofMethod,rplColor), hBL, rr, gg, bb, aa); //bottom row
\r
623 AddAveragingCont(GetPixelColorWithOverflow(x,yi2,ofMethod,rplColor), hTR, rr, gg, bb, aa); //top row
\r
625 //leftmost and rightmost column
\r
626 for (y=yi1+1; y<yi2; y++) {
\r
627 AddAveragingCont(GetPixelColorWithOverflow(xi1,y,ofMethod,rplColor), wBL, rr, gg, bb, aa); //left column
\r
628 AddAveragingCont(GetPixelColorWithOverflow(xi2,y,ofMethod,rplColor), wTR, rr, gg, bb, aa); //right column
\r
630 for (y=yi1+1; y<yi2; y++) {
\r
631 for (x=xi1+1; x<xi2; x++) {
\r
632 color=GetPixelColorWithOverflow(x,y,ofMethod,rplColor);
\r
634 gg+=color.rgbGreen;
\r
636 #if CXIMAGE_SUPPORT_ALPHA
\r
637 aa+=color.rgbReserved;
\r
642 //width or height greater than one:
\r
643 CxRect2 intersect; //intersection with current pixel
\r
645 for (y=yi1; y<=yi2; y++) {
\r
646 for (x=xi1; x<=xi2; x++) {
\r
647 intersect=area.CrossSection(CxRect2(((float)x)-0.5f, ((float)y)-0.5f, ((float)x)+0.5f, ((float)y)+0.5f));
\r
648 center=intersect.Center();
\r
649 color=GetPixelColorInterpolated(center.x, center.y, inMethod, ofMethod, rplColor);
\r
650 cps=intersect.Surface();
\r
651 rr+=color.rgbRed*cps;
\r
652 gg+=color.rgbGreen*cps;
\r
653 bb+=color.rgbBlue*cps;
\r
654 #if CXIMAGE_SUPPORT_ALPHA
\r
655 aa+=color.rgbReserved*cps;
\r
662 rr/=s; gg/=s; bb/=s; aa/=s;
\r
663 if (rr>255) rr=255; if (rr<0) rr=0; color.rgbRed=(BYTE) rr;
\r
664 if (gg>255) gg=255; if (gg<0) gg=0; color.rgbGreen=(BYTE) gg;
\r
665 if (bb>255) bb=255; if (bb<0) bb=0; color.rgbBlue=(BYTE) bb;
\r
666 #if CXIMAGE_SUPPORT_ALPHA
\r
667 if (AlphaIsValid()) {
\r
668 if (aa>255) aa=255; if (aa<0) aa=0; color.rgbReserved=(BYTE) aa;
\r
675 ////////////////////////////////////////////////////////////////////////////////
\r
676 float CxImage::KernelBSpline(const float x)
\r
678 if (x>2.0f) return 0.0f;
\r
679 // thanks to Kristian Kratzenstein
\r
681 float xm1 = x - 1.0f; // Was calculatet anyway cause the "if((x-1.0f) < 0)"
\r
682 float xp1 = x + 1.0f;
\r
683 float xp2 = x + 2.0f;
\r
685 if ((xp2) <= 0.0f) a = 0.0f; else a = xp2*xp2*xp2; // Only float, not float -> double -> float
\r
686 if ((xp1) <= 0.0f) b = 0.0f; else b = xp1*xp1*xp1;
\r
687 if (x <= 0) c = 0.0f; else c = x*x*x;
\r
688 if ((xm1) <= 0.0f) d = 0.0f; else d = xm1*xm1*xm1;
\r
690 return (0.16666666666666666667f * (a - (4.0f * b) + (6.0f * c) - (4.0f * d)));
\r
692 /* equivalent <VladimÃr Kloucek>
\r
696 return((2.0f+x)*(2.0f+x)*(2.0f+x)*0.16666666666666666667f);
\r
698 return((4.0f+x*x*(-6.0f-3.0f*x))*0.16666666666666666667f);
\r
700 return((4.0f+x*x*(-6.0f+3.0f*x))*0.16666666666666666667f);
\r
702 return((2.0f-x)*(2.0f-x)*(2.0f-x)*0.16666666666666666667f);
\r
707 ////////////////////////////////////////////////////////////////////////////////
\r
709 * Bilinear interpolation kernel:
\r
712 | 1-t , if 0 <= t <= 1
\r
713 h(t) = | t+1 , if -1 <= t < 0
\r
719 float CxImage::KernelLinear(const float t)
\r
721 // if (0<=t && t<=1) return 1-t;
\r
722 // if (-1<=t && t<0) return 1+t;
\r
725 //<VladimÃr Kloucek>
\r
735 ////////////////////////////////////////////////////////////////////////////////
\r
737 * Bicubic interpolation kernel (a=-1):
\r
740 | 1-2|t|**2+|t|**3 , if |t| < 1
\r
741 h(t) = | 4-8|t|+5|t|**2-|t|**3 , if 1<=|t|<2
\r
747 float CxImage::KernelCubic(const float t)
\r
749 float abs_t = (float)fabs(t);
\r
750 float abs_t_sq = abs_t * abs_t;
\r
751 if (abs_t<1) return 1-2*abs_t_sq+abs_t_sq*abs_t;
\r
752 if (abs_t<2) return 4 - 8*abs_t +5*abs_t_sq - abs_t_sq*abs_t;
\r
756 ////////////////////////////////////////////////////////////////////////////////
\r
758 * Bicubic kernel (for a=-1 it is the same as BicubicKernel):
\r
761 | (a+2)|t|**3 - (a+3)|t|**2 + 1 , |t| <= 1
\r
762 h(t) = | a|t|**3 - 5a|t|**2 + 8a|t| - 4a , 1 < |t| <= 2
\r
766 * Often used values for a are -1 and -1/2.
\r
768 float CxImage::KernelGeneralizedCubic(const float t, const float a)
\r
770 float abs_t = (float)fabs(t);
\r
771 float abs_t_sq = abs_t * abs_t;
\r
772 if (abs_t<1) return (a+2)*abs_t_sq*abs_t - (a+3)*abs_t_sq + 1;
\r
773 if (abs_t<2) return a*abs_t_sq*abs_t - 5*a*abs_t_sq + 8*a*abs_t - 4*a;
\r
777 ////////////////////////////////////////////////////////////////////////////////
\r
779 * Lanczos windowed sinc interpolation kernel with radius r.
\r
782 h(t) = | sinc(t)*sinc(t/r) , if |t|<r
\r
788 float CxImage::KernelLanczosSinc(const float t, const float r)
\r
790 if (fabs(t) > r) return 0;
\r
791 if (t==0) return 1;
\r
794 return (float)((sin(pit)/pit) * (sin(pitd)/pitd));
\r
797 ////////////////////////////////////////////////////////////////////////////////
\r
798 float CxImage::KernelBox(const float x)
\r
806 ////////////////////////////////////////////////////////////////////////////////
\r
807 float CxImage::KernelHermite(const float x)
\r
812 return (-2.0f*x-3.0f)*x*x+1.0f;
\r
814 return (2.0f*x-3.0f)*x*x+1.0f;
\r
816 // if (fabs(x)>1) return 0.0f;
\r
817 // return(0.5f+0.5f*(float)cos(PI*x));
\r
819 ////////////////////////////////////////////////////////////////////////////////
\r
820 float CxImage::KernelHanning(const float x)
\r
822 if (fabs(x)>1) return 0.0f;
\r
823 return (0.5f+0.5f*(float)cos(PI*x))*((float)sin(PI*x)/(PI*x));
\r
825 ////////////////////////////////////////////////////////////////////////////////
\r
826 float CxImage::KernelHamming(const float x)
\r
831 return 0.92f*(-2.0f*x-3.0f)*x*x+1.0f;
\r
833 return 0.92f*(2.0f*x-3.0f)*x*x+1.0f;
\r
835 // if (fabs(x)>1) return 0.0f;
\r
836 // return(0.54f+0.46f*(float)cos(PI*x));
\r
838 ////////////////////////////////////////////////////////////////////////////////
\r
839 float CxImage::KernelSinc(const float x)
\r
843 return((float)sin(PI*x)/(PI*x));
\r
845 ////////////////////////////////////////////////////////////////////////////////
\r
846 float CxImage::KernelBlackman(const float x)
\r
848 //if (fabs(x)>1) return 0.0f;
\r
849 return (0.42f+0.5f*(float)cos(PI*x)+0.08f*(float)cos(2.0f*PI*x));
\r
851 ////////////////////////////////////////////////////////////////////////////////
\r
852 float CxImage::KernelBessel_J1(const float x)
\r
858 static const double
\r
861 0.581199354001606143928050809e+21,
\r
862 -0.6672106568924916298020941484e+20,
\r
863 0.2316433580634002297931815435e+19,
\r
864 -0.3588817569910106050743641413e+17,
\r
865 0.2908795263834775409737601689e+15,
\r
866 -0.1322983480332126453125473247e+13,
\r
867 0.3413234182301700539091292655e+10,
\r
868 -0.4695753530642995859767162166e+7,
\r
869 0.270112271089232341485679099e+4
\r
873 0.11623987080032122878585294e+22,
\r
874 0.1185770712190320999837113348e+20,
\r
875 0.6092061398917521746105196863e+17,
\r
876 0.2081661221307607351240184229e+15,
\r
877 0.5243710262167649715406728642e+12,
\r
878 0.1013863514358673989967045588e+10,
\r
879 0.1501793594998585505921097578e+7,
\r
880 0.1606931573481487801970916749e+4,
\r
886 for (i=7; i >= 0; i--)
\r
891 return (float)(p/q);
\r
893 ////////////////////////////////////////////////////////////////////////////////
\r
894 float CxImage::KernelBessel_P1(const float x)
\r
900 static const double
\r
903 0.352246649133679798341724373e+5,
\r
904 0.62758845247161281269005675e+5,
\r
905 0.313539631109159574238669888e+5,
\r
906 0.49854832060594338434500455e+4,
\r
907 0.2111529182853962382105718e+3,
\r
908 0.12571716929145341558495e+1
\r
912 0.352246649133679798068390431e+5,
\r
913 0.626943469593560511888833731e+5,
\r
914 0.312404063819041039923015703e+5,
\r
915 0.4930396490181088979386097e+4,
\r
916 0.2030775189134759322293574e+3,
\r
922 for (i=4; i >= 0; i--)
\r
924 p = p*(8.0/x)*(8.0/x)+Pone[i];
\r
925 q = q*(8.0/x)*(8.0/x)+Qone[i];
\r
927 return (float)(p/q);
\r
929 ////////////////////////////////////////////////////////////////////////////////
\r
930 float CxImage::KernelBessel_Q1(const float x)
\r
936 static const double
\r
939 0.3511751914303552822533318e+3,
\r
940 0.7210391804904475039280863e+3,
\r
941 0.4259873011654442389886993e+3,
\r
942 0.831898957673850827325226e+2,
\r
943 0.45681716295512267064405e+1,
\r
944 0.3532840052740123642735e-1
\r
948 0.74917374171809127714519505e+4,
\r
949 0.154141773392650970499848051e+5,
\r
950 0.91522317015169922705904727e+4,
\r
951 0.18111867005523513506724158e+4,
\r
952 0.1038187585462133728776636e+3,
\r
958 for (i=4; i >= 0; i--)
\r
960 p = p*(8.0/x)*(8.0/x)+Pone[i];
\r
961 q = q*(8.0/x)*(8.0/x)+Qone[i];
\r
963 return (float)(p/q);
\r
965 ////////////////////////////////////////////////////////////////////////////////
\r
966 float CxImage::KernelBessel_Order1(float x)
\r
976 return(p*KernelBessel_J1(x));
\r
977 q = (float)sqrt(2.0f/(PI*x))*(float)(KernelBessel_P1(x)*(1.0f/sqrt(2.0f)*(sin(x)-cos(x)))-8.0f/x*KernelBessel_Q1(x)*
\r
978 (-1.0f/sqrt(2.0f)*(sin(x)+cos(x))));
\r
983 ////////////////////////////////////////////////////////////////////////////////
\r
984 float CxImage::KernelBessel(const float x)
\r
988 return(KernelBessel_Order1(PI*x)/(2.0f*x));
\r
990 ////////////////////////////////////////////////////////////////////////////////
\r
991 float CxImage::KernelGaussian(const float x)
\r
993 return (float)(exp(-2.0f*x*x)*0.79788456080287f/*sqrt(2.0f/PI)*/);
\r
995 ////////////////////////////////////////////////////////////////////////////////
\r
996 float CxImage::KernelQuadratic(const float x)
\r
1001 return(0.5f*(x+1.5f)*(x+1.5f));
\r
1003 return(0.75f-x*x);
\r
1005 return(0.5f*(x-1.5f)*(x-1.5f));
\r
1008 ////////////////////////////////////////////////////////////////////////////////
\r
1009 float CxImage::KernelMitchell(const float x)
\r
1011 #define KM_B (1.0f/3.0f)
\r
1012 #define KM_C (1.0f/3.0f)
\r
1013 #define KM_P0 (( 6.0f - 2.0f * KM_B ) / 6.0f)
\r
1014 #define KM_P2 ((-18.0f + 12.0f * KM_B + 6.0f * KM_C) / 6.0f)
\r
1015 #define KM_P3 (( 12.0f - 9.0f * KM_B - 6.0f * KM_C) / 6.0f)
\r
1016 #define KM_Q0 (( 8.0f * KM_B + 24.0f * KM_C) / 6.0f)
\r
1017 #define KM_Q1 ((-12.0f * KM_B - 48.0f * KM_C) / 6.0f)
\r
1018 #define KM_Q2 (( 6.0f * KM_B + 30.0f * KM_C) / 6.0f)
\r
1019 #define KM_Q3 (( -1.0f * KM_B - 6.0f * KM_C) / 6.0f)
\r
1024 return(KM_Q0-x*(KM_Q1-x*(KM_Q2-x*KM_Q3)));
\r
1026 return(KM_P0+x*x*(KM_P2-x*KM_P3));
\r
1028 return(KM_P0+x*x*(KM_P2+x*KM_P3));
\r
1030 return(KM_Q0+x*(KM_Q1+x*(KM_Q2+x*KM_Q3)));
\r
1033 ////////////////////////////////////////////////////////////////////////////////
\r
1034 float CxImage::KernelCatrom(const float x)
\r
1039 return(0.5f*(4.0f+x*(8.0f+x*(5.0f+x))));
\r
1041 return(0.5f*(2.0f+x*x*(-5.0f-3.0f*x)));
\r
1043 return(0.5f*(2.0f+x*x*(-5.0f+3.0f*x)));
\r
1045 return(0.5f*(4.0f+x*(-8.0f+x*(5.0f-x))));
\r
1048 ////////////////////////////////////////////////////////////////////////////////
\r
1049 float CxImage::KernelPower(const float x, const float a)
\r
1051 if (fabs(x)>1) return 0.0f;
\r
1052 return (1.0f - (float)fabs(pow(x,a)));
\r
1054 ////////////////////////////////////////////////////////////////////////////////
\r