| 18 |
|
* calculation, we sort the data into 3 half-planes and |
| 19 |
|
* perform simple tests to see which neighbor is closest in |
| 20 |
|
* a each direction. Once we have our approximate neighborhood |
| 21 |
< |
* for a sample, we can use it in a Gaussian weighting scheme |
| 22 |
< |
* with anisotropic surround. This gives us a fairly smooth |
| 23 |
< |
* interpolation however the sample points may be initially |
| 24 |
< |
* distributed. Evaluation is accelerated by use of a fast |
| 25 |
< |
* approximation to the atan2(y,x) function. |
| 21 |
> |
* for a sample, we can use it in a modified Gaussian weighting |
| 22 |
> |
* scheme with anisotropic surround. Harmonic weighting is added |
| 23 |
> |
* to reduce the influence of distant neighbors. This yields a |
| 24 |
> |
* smooth interpolation regardless of how the sample points are |
| 25 |
> |
* initiallydistributed. Evaluation is accelerated by use of a |
| 26 |
> |
* fast approximation to the atan2(y,x) function. |
| 27 |
|
**************************************************************/ |
| 28 |
|
|
| 29 |
|
#include <stdio.h> |
| 183 |
|
sort_samples(sortord, ip, ang); |
| 184 |
|
/* find nearest neighbors each side */ |
| 185 |
|
for (i = ip->ns; i--; ) { |
| 186 |
< |
const int rpos = rightrndx[sortord[i].si]; |
| 186 |
< |
const int lpos = leftrndx[sortord[i].si]; |
| 186 |
> |
const int ii = sortord[i].si; |
| 187 |
|
int j; |
| 188 |
< |
/* preload with large radius */ |
| 189 |
< |
ip->ra[i][bd] = ip->ra[i][bd+NI2DIR/2] = encode_radius(ip, |
| 188 |
> |
/* preload with large radii */ |
| 189 |
> |
ip->ra[ii][bd] = ip->ra[ii][bd+NI2DIR/2] = encode_radius(ip, |
| 190 |
|
.25*(sortord[ip->ns-1].dm - sortord[0].dm)); |
| 191 |
|
for (j = i; ++j < ip->ns; ) /* nearest above */ |
| 192 |
< |
if (rightrndx[sortord[j].si] > rpos && |
| 193 |
< |
leftrndx[sortord[j].si] < lpos) { |
| 194 |
< |
ip->ra[i][bd] = encode_radius(ip, |
| 192 |
> |
if (rightrndx[sortord[j].si] > rightrndx[ii] && |
| 193 |
> |
leftrndx[sortord[j].si] < leftrndx[ii]) { |
| 194 |
> |
ip->ra[ii][bd] = encode_radius(ip, |
| 195 |
|
.5*(sortord[j].dm - sortord[i].dm)); |
| 196 |
|
break; |
| 197 |
|
} |
| 198 |
|
for (j = i; j-- > 0; ) /* nearest below */ |
| 199 |
< |
if (rightrndx[sortord[j].si] < rpos && |
| 200 |
< |
leftrndx[sortord[j].si] > lpos) { |
| 201 |
< |
ip->ra[i][bd+NI2DIR/2] = encode_radius(ip, |
| 199 |
> |
if (rightrndx[sortord[j].si] < rightrndx[ii] && |
| 200 |
> |
leftrndx[sortord[j].si] > leftrndx[ii]) { |
| 201 |
> |
ip->ra[ii][bd+NI2DIR/2] = encode_radius(ip, |
| 202 |
|
.5*(sortord[i].dm - sortord[j].dm)); |
| 203 |
|
break; |
| 204 |
|
} |
| 211 |
|
return(1); |
| 212 |
|
} |
| 213 |
|
|
| 214 |
< |
/* private call returns log of raw weight for a particular sample */ |
| 214 |
> |
/* private call returns raw weight for a particular sample */ |
| 215 |
|
static double |
| 216 |
< |
get_ln_wt(const INTERP2 *ip, const int i, double x, double y) |
| 216 |
> |
get_wt(const INTERP2 *ip, const int i, double x, double y) |
| 217 |
|
{ |
| 218 |
< |
double dir, rd; |
| 218 |
> |
double dir, rd, d2; |
| 219 |
|
int ri; |
| 220 |
|
/* get relative direction */ |
| 221 |
|
x -= ip->spt[i][0]; |
| 228 |
|
rd -= (double)ri; |
| 229 |
|
rd = (1.-rd)*ip->ra[i][ri] + rd*ip->ra[i][(ri+1)%NI2DIR]; |
| 230 |
|
rd = ip->smf * DECODE_RAD(ip, rd); |
| 231 |
< |
/* return log of Gaussian weight */ |
| 232 |
< |
return( (x*x + y*y) / (-2.*rd*rd) ); |
| 231 |
> |
d2 = x*x + y*y; |
| 232 |
> |
/* Gaussian times harmonic weighting */ |
| 233 |
> |
return( exp(d2/(-2.*rd*rd)) * ip->rmin/(ip->rmin + sqrt(d2)) ); |
| 234 |
|
} |
| 235 |
|
|
| 236 |
|
/* Assign full set of normalized weights to interpolate the given position */ |
| 248 |
|
|
| 249 |
|
wnorm = 0; /* compute raw weights */ |
| 250 |
|
for (i = ip->ns; i--; ) { |
| 251 |
< |
double wt = get_ln_wt(ip, i, x, y); |
| 251 |
< |
if (wt < -21.) { |
| 252 |
< |
wtv[i] = 0; /* ignore weights < 1e-9 */ |
| 253 |
< |
continue; |
| 254 |
< |
} |
| 255 |
< |
wt = exp(wt); /* Gaussian weight */ |
| 251 |
> |
double wt = get_wt(ip, i, x, y); |
| 252 |
|
wtv[i] = wt; |
| 253 |
|
wnorm += wt; |
| 254 |
|
} |
| 276 |
|
return(0); |
| 277 |
|
/* identify top n weights */ |
| 278 |
|
for (i = ip->ns; i--; ) { |
| 279 |
< |
const double lnwt = get_ln_wt(ip, i, x, y); |
| 279 |
> |
const double wti = get_wt(ip, i, x, y); |
| 280 |
|
for (j = nn; j > 0; j--) { |
| 281 |
< |
if (wt[j-1] >= lnwt) |
| 281 |
> |
if (wt[j-1] >= wti) |
| 282 |
|
break; |
| 283 |
|
if (j < n) { |
| 284 |
|
wt[j] = wt[j-1]; |
| 286 |
|
} |
| 287 |
|
} |
| 288 |
|
if (j < n) { /* add/insert sample */ |
| 289 |
< |
wt[j] = lnwt; |
| 289 |
> |
wt[j] = wti; |
| 290 |
|
si[j] = i; |
| 291 |
|
nn += (nn < n); |
| 292 |
|
} |
| 293 |
|
} |
| 294 |
< |
wnorm = 0; /* exponentiate and normalize */ |
| 295 |
< |
for (j = nn; j--; ) { |
| 296 |
< |
double dwt = exp(wt[j]); |
| 301 |
< |
wt[j] = dwt; |
| 302 |
< |
wnorm += dwt; |
| 303 |
< |
} |
| 294 |
> |
wnorm = 0; /* normalize sample weights */ |
| 295 |
> |
for (j = nn; j--; ) |
| 296 |
> |
wnorm += wt[j]; |
| 297 |
|
if (wnorm <= 0) |
| 298 |
|
return(0); |
| 299 |
|
wnorm = 1./wnorm; |
