CFA line denoise improvements on behalf of Emil and Jacques; see issue #408

This commit is contained in:
Oliver Duis 2010-12-21 22:54:10 +01:00
parent 84b7e9b6ea
commit f90f4fb7d5

View File

@ -27,8 +27,8 @@
#define CLASS
/*#include <ctype.h>
/*
#include <ctype.h>
#include <errno.h>
#include <fcntl.h>
#include <float.h>
@ -59,22 +59,24 @@ void RawImageSource::CLASS cfa_linedn(float noise)
{
// local variables
int height=H, width=W;
int top, left, row, col;
int rr, cc, indx, i, j;
int ex, ey;
int verbose=1;
const float clip_pt = 0.8*initialGain* 65535.0;
float eps=1e-5; //tolerance to avoid dividing by zero
float gauss[5] = {0.20416368871516755, 0.18017382291138087, 0.1238315368057753, 0.0662822452863612, 0.02763055063889883};
float rolloff[8] = {0, 0.135335, 0.249352, 0.411112, 0.606531, 0.800737, 0.945959, 1}; //gaussian with sigma=3
float window[8] = {0, .25, .75, 1, 1, .75, .25, 0}; //sine squared
float noisevar, linehvar, linevvar, coeffsq;
const float gauss[5] = {0.20416368871516755, 0.18017382291138087, 0.1238315368057753, 0.0662822452863612, 0.02763055063889883};
const float rolloff[8] = {0, 0.135335, 0.249352, 0.411112, 0.606531, 0.800737, 0.945959, 1}; //gaussian with sigma=3
const float window[8] = {0, .25, .75, 1, 1, .75, .25, 0}; //sine squared
float aarr[4][8][8], *bbrr[4][8], **dctblock[4];
for (int j=0; j<4; j++) {
for (int i = 0; i < 8; i++)
bbrr[j][i] = aarr[j][i];
dctblock[j] = bbrr[j];
}
float aarr[8][8], *dctblock[8];
for (i = 0; i < 8; i++) dctblock[i] = aarr[i];
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
@ -84,124 +86,193 @@ void RawImageSource::CLASS cfa_linedn(float noise)
}
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
float noisevar=SQR(3*noise*65535); // _noise_ (as a fraction of saturation) is input to the algorithm
volatile double progress = 0.0;
#pragma omp parallel
{
noisevar=SQR(3*noise*65535); // _noise_ (as a fraction of saturation) is input to the algorithm
float *cfain= new float[TS*TS];
float *cfablur= new float[TS*TS];
float *cfadiff= new float[TS*TS];
float *cfadn= new float[TS*TS];
/*
char *buffer; // TS*TS*16
float (*cfain); // TS*TS*4
float (*cfablur); // TS*TS*4
float (*cfadiff); // TS*TS*4
float (*cfadn); // TS*TS*4
buffer = (char *) calloc((3*sizeof(float)+sizeof(int))*TS*TS,1);// 16
// merror(buffer,"cfa_linedn()");
// memset(buffer,0,16*TS*TS);
cfain = (float (*)) buffer; //pointers to rows of array
cfablur = (float (*)) (buffer + sizeof(float)*TS*TS); //4
cfadiff = (float (*)) (buffer + 2*sizeof(float)*TS*TS);//8
cfadn = (float (*)) (buffer + 3*sizeof(float)*TS*TS);//12
*/
// Main algorithm: Tile loop
for (top=0; top < height-16; top += TS-32)
for (left=0; left < width-16; left += TS-32) {
#pragma omp for schedule(dynamic) nowait
for (int top=0; top < height-16; top += TS-32)
for (int left=0; left < width-16; left += TS-32) {
int bottom = MIN( top+TS,height);
int right = MIN(left+TS, width);
int numrows = bottom - top;
int numcols = right - left;
int indx1;
// load CFA data; data should be in linear gamma space, before white balance multipliers are applied
for (rr=top; rr < top+numrows; rr++)
for (cc=left, indx=(rr-top)*TS; cc < left+numcols; cc++, indx++) {
for (int rr=top; rr < top+numrows; rr++)
for (int cc=left, indx=(rr-top)*TS; cc < left+numcols; cc++, indx++) {
cfain[indx] = rawData[rr][cc];
}
//pad the block to a multiple of 16 on both sides
if (numcols < TS) {
indx=numcols % 16;
for (i=0; i<(16-indx); i++)
for (rr=0; rr<numrows; rr++)
indx1=numcols % 16;
for (int i=0; i<(16-indx1); i++)
for (int rr=0; rr<numrows; rr++)
cfain[(rr)*TS+numcols+i+1]=cfain[(rr)*TS+numcols-i];
numcols += 16-indx;
numcols += 16-indx1;
}
if (numrows < TS) {
indx=numrows % 16;
for (i=0; i<(16-indx); i++)
for (cc=0; cc<numcols; cc++)
indx1=numrows % 16;
for (int i=0; i<(16-indx1); i++)
for (int cc=0; cc<numcols; cc++)
cfain[(numrows+i+1)*TS+cc]=cfain[(numrows-i)*TS+cc];
numrows += 16-indx;
numrows += 16-indx1;
}
//The cleaning algorithm starts here
//gaussian blur of CFA data
for (rr=8; rr < numrows-8; rr++)
for (indx=rr*TS; indx < rr*TS+numcols; indx++) {
for (int rr=8; rr < numrows-8; rr++)
for (int indx=rr*TS; indx < rr*TS+numcols; indx++) {
cfablur[indx]=gauss[0]*cfain[indx];
for (i=1; i<5; i++) {
for (int i=1; i<5; i++) {
cfablur[indx] += gauss[i]*(cfain[indx-(2*i)*TS]+cfain[indx+(2*i)*TS]);
}
}
for (rr=8; rr < numrows-8; rr++)
for (indx=rr*TS+8; indx < rr*TS+numcols-8; indx++) {
for (int rr=8; rr < numrows-8; rr++)
for (int indx=rr*TS+8; indx < rr*TS+numcols-8; indx++) {
cfadn[indx] = gauss[0]*cfablur[indx];
for (i=1; i<5; i++) {
for (int i=1; i<5; i++) {
cfadn[indx] += gauss[i]*(cfablur[indx-2*i]+cfablur[indx+2*i]);
}
cfadiff[indx]=cfain[indx]-cfadn[indx]; // hipass cfa data
}
//begin block DCT
for (ey=0; ey<2; ey++) // (ex,ey) specify RGGB subarray
for (ex=0; ex<2; ex++)
for (rr=8+ey; rr < numrows-22; rr+=8) // (rr,cc) shift by 8 to overlap blocks
for (cc=8+ex; cc < numcols-22; cc+=8) {
float linehvar[4], linevvar[4], noisefactor[4][8][2], coeffsq;
#pragma omp critical
{
for (int rr=8; rr < numrows-22; rr+=8) // (rr,cc) shift by 8 to overlap blocks
for (int cc=8; cc < numcols-22; cc+=8) {
for (int ey=0; ey<2; ey++) // (ex,ey) specify RGGB subarray
for (int ex=0; ex<2; ex++) {
//grab an 8x8 block of a given RGGB channel
for (i=0; i<8; i++)
for (j=0; j<8; j++) {
dctblock[i][j]=cfadiff[(rr+2*i)*TS+cc+2*j];
for (int i=0; i<8; i++)
for (int j=0; j<8; j++) {
dctblock[2*ey+ex][i][j]=cfadiff[(rr+2*i+ey)*TS+cc+2*j+ex];
}
ddct8x8s(-1, dctblock); //forward DCT
ddct8x8s(-1, dctblock[2*ey+ex]); //forward DCT
}
for (int ey=0; ey<2; ey++) // (ex,ey) specify RGGB subarray
for (int ex=0; ex<2; ex++) {
linehvar[2*ey+ex]=linevvar[2*ey+ex]=0;
for (int i=4; i<8; i++) {
linehvar[2*ey+ex] += SQR(dctblock[2*ey+ex][0][i]);
linevvar[2*ey+ex] += SQR(dctblock[2*ey+ex][i][0]);
linehvar=linevvar=0;
for (i=4; i<8; i++) {
linehvar += SQR(dctblock[0][i]);
linevvar += SQR(dctblock[i][0]);
}
//Wiener filter for line denoising; roll off low frequencies
if (noisevar>linehvar) {
for (i=1; i<8; i++) {
coeffsq=SQR(dctblock[0][i]);
dctblock[0][i] *= coeffsq/(coeffsq+rolloff[i]*noisevar+eps);
for (int i=1; i<8; i++) {
coeffsq = SQR(dctblock[2*ey+ex][i][0]);//vertical
noisefactor[2*ey+ex][i][0] = coeffsq/(coeffsq+rolloff[i]*noisevar+eps);
coeffsq = SQR(dctblock[2*ey+ex][0][i]);//horizontal
noisefactor[2*ey+ex][i][1] = coeffsq/(coeffsq+rolloff[i]*noisevar+eps);
// noisefactor labels are [RGGB subarray][row/col position][0=vert,1=hor]
}
}
if (noisevar>linevvar) {
for (i=1; i<8; i++) {
coeffsq=SQR(dctblock[i][0]);
dctblock[i][0] *= coeffsq/(coeffsq+rolloff[i]*noisevar+eps);
//horizontal lines
if (4*noisevar>(linehvar[0]+linehvar[1])) {//horizontal lines
for (int i=1; i<8; i++) {
dctblock[0][0][i] *= 0.5*(noisefactor[0][i][1]+noisefactor[1][i][1]);//or should we use MIN???
dctblock[1][0][i] *= 0.5*(noisefactor[0][i][1]+noisefactor[1][i][1]);//or should we use MIN???
}
}
if (4*noisevar>(linehvar[2]+linehvar[3])) {//horizontal lines
for (int i=1; i<8; i++) {
dctblock[2][0][i] *= 0.5*(noisefactor[2][i][1]+noisefactor[3][i][1]);//or should we use MIN???
dctblock[3][0][i] *= 0.5*(noisefactor[2][i][1]+noisefactor[3][i][1]);//or should we use MIN???
}
}
ddct8x8s(1, dctblock); //inverse DCT
//vertical lines
if (4*noisevar>(linevvar[0]+linevvar[2])) {//vertical lines
for (int i=1; i<8; i++) {
dctblock[0][i][0] *= 0.5*(noisefactor[0][i][0]+noisefactor[2][i][0]);//or should we use MIN???
dctblock[2][i][0] *= 0.5*(noisefactor[0][i][0]+noisefactor[2][i][0]);//or should we use MIN???
}
}
if (4*noisevar>(linevvar[1]+linevvar[3])) {//vertical lines
for (int i=1; i<8; i++) {
dctblock[1][i][0] *= 0.5*(noisefactor[1][i][0]+noisefactor[3][i][0]);//or should we use MIN???
dctblock[3][i][0] *= 0.5*(noisefactor[1][i][0]+noisefactor[3][i][0]);//or should we use MIN???
}
}
for (int ey=0; ey<2; ey++) // (ex,ey) specify RGGB subarray
for (int ex=0; ex<2; ex++) {
ddct8x8s(1, dctblock[2*ey+ex]); //inverse DCT
//multiply by window fn and add to output (cfadn)
for (i=0; i<8; i++)
for (j=0; j<8; j++) {
cfadn[(rr+2*i)*TS+cc+2*j] += window[i]*window[j]*dctblock[i][j];
}
}
for (int i=0; i<8; i++)
for (int j=0; j<8; j++) {
cfadn[(rr+2*i+ey)*TS+cc+2*j+ex] += window[i]*window[j]*dctblock[2*ey+ex][i][j];
}
}
}
}
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// copy smoothed results back to image matrix
for (rr=16; rr < numrows-16; rr++) {
row = rr + top;
for (col=16+left, indx=rr*TS+16; indx < rr*TS+numcols-16; indx++, col++) {
for (int rr=16; rr < numrows-16; rr++) {
int row = rr + top;
for (int col=16+left, indx=rr*TS+16; indx < rr*TS+numcols-16; indx++, col++) {
if (rawData[row][col]<clip_pt && cfadn[indx]<clip_pt)
rawData[row][col] = CLIP((int)(cfadn[indx]+ 0.5));
}
}
if(plistener) plistener->setProgress(fabs((float)top/height));
progress+=(double)((TS-32)*(TS-32))/(height*width);
if (progress>1.0)
{
progress=1.0;
}
if(plistener) plistener->setProgress(progress);
//if(plistener) plistener->setProgress(fabs((float)top/height));
}
// clean up
delete [] cfain;
delete [] cfablur;
delete [] cfadiff;
delete [] cfadn;
//free(buffer);
}
}
#undef TS