问题
how to scale and by which factor to scale dctmtx coefficients from float to get following integer values:
float dctmtx:
( (0.3536 0.3536 0.3536 0.3536 0.3536 0.3536 0.3536 0.3536),
(0.4904 0.4157 0.2778 0.0975 -0.0975 -0.2778 -0.4157 -0.4904),
(0.4619 0.1913 -0.1913 -0.4619 -0.4619 -0.1913 0.1913 0.4619),
(0.4157 -0.0975 -0.4904 -0.2778 0.2778 0.4904 0.0975 -0.4157),
(0.3536 -0.3536 -0.3536 0.3536 0.3536 -0.3536 -0.3536 0.3536),
(0.2778 -0.4904 0.0975 0.4157 -0.4157 -0.0975 0.4904 -0.2778),
(0.1913 -0.4619 0.4619 -0.1913 -0.1913 0.4619 -0.4619 0.1913),
(0.0975 -0.2778 0.4157 -0.4904 0.4904 -0.4157 0.2778 -0.0975)
)
integer dctmtx:
(( 125, 122, 115, 103, 88, 69, 47, 24 ),
( 125, 103, 47, -24, -88, -122, -115, -69 ),
( 125, 69, -47, -122, -88, 24, 115, 103 ),
( 125, 24, -115, -69, 88, 103, -47, -122 ),
( 125, -24, -115, 69, 88, -103, -47, 122 ),
( 125, -69, -47, 122, -88, -24, 115, -103 ),
( 125, -103, 47, 24, -88, 122, -115, 69 ),
( 125, -122, 115, -103, 88, -69, 47, -24 )
);
回答1:
If you read up on the discrete cosine transform, you will find that the basic coefficient is
cos(pi*i*(2*j+1)/16), i,j=0..7
Then the first table consists of these values scaled by 0.5, except for the first row/column, which is scaled by 0.25*sqrt(2)=1/sqrt(8). Which is the correct way to obtain an orthogonal matrix. The square sum of the first column is 8, of the others 4.
The second table is the rounded results when multiplying the cosine values with 125, uniformly. Here one has to take care to properly rescale the vector when using the transpose matrix to compute the inverse transform.
First table reproduced, except for the first column:
> [[ Cos(pi*i*(2*j+1)/16)/2 : i in [0..7] ]: j in [0..7] ];
[
[ 0.5, 0.49039264, 0.46193977, 0.41573481, 0.35355339, 0.27778512, 0.19134172, 0.09754516 ],
[ 0.5, 0.41573481, 0.19134172, -0.09754516, -0.35355339, -0.49039264, -0.46193977, -0.27778512 ],
[ 0.5, 0.27778512, -0.19134172, -0.49039264, -0.35355339, 0.09754516, 0.46193977, 0.41573481 ],
[ 0.5, 0.09754516, -0.46193977, -0.27778512, 0.35355339, 0.41573481, -0.19134172, -0.49039264 ],
[ 0.5, -0.09754516, -0.46193977, 0.27778512, 0.35355339, -0.41573481, -0.19134172, 0.49039264 ],
[ 0.5, -0.27778512, -0.19134172, 0.49039264, -0.35355339, -0.09754516, 0.46193977, -0.41573481 ],
[ 0.5, -0.41573481, 0.19134172, 0.09754516, -0.35355339, 0.49039264, -0.46193977, 0.27778512 ],
[ 0.5, -0.49039264, 0.46193977, -0.41573481, 0.35355339, -0.27778512, 0.19134172, -0.09754516 ]
]
Second table, before integer rounding
> [[ Cos( pi*i*(2*j+1)/16 ) *125 : i in [0..7] ]: j in [0..7] ];
[
[ 125, 122.5982, 115.4849, 103.9337, 88.3883, 69.4463, 47.8354, 24.3863 ],
[ 125, 103.9337, 47.8354, -24.3863, -88.3883, -122.5982, -115.4849, -69.4463 ],
[ 125, 69.4463, -47.8354, -122.5982, -88.3883, 24.3863, 115.4849, 103.9337 ],
[ 125, 24.3863, -115.4849, -69.4463, 88.3883, 103.9337, -47.8354, -122.5982 ],
[ 125, -24.3863, -115.4849, 69.4463, 88.3883, -103.9337, -47.8354, 122.5982 ],
[ 125, -69.4463, -47.8354, 122.5982, -88.3883, -24.3863, 115.4849, -103.9337 ],
[ 125, -103.9337, 47.8354, 24.3863, -88.3883, 122.5982, -115.4849, 69.4463 ],
[ 125, -122.5982, 115.4849, -103.9337, 88.3883, -69.4463, 47.8354, -24.3863 ]
]
回答2:
Besides one of the two matrices being rotated the two don't appear to have a direct linear relationship:
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
int main (int argc, char *argv[])
{
float dctmtx[8][8] =
{ 0.3536, 0.3536, 0.3536, 0.3536, 0.3536, 0.3536, 0.3536, 0.3536,
0.4904, 0.4157, 0.2778, 0.0975, -0.0975, -0.2778, -0.4157, -0.4904,
0.4619, 0.1913, -0.1913, -0.4619, -0.4619, -0.1913, 0.1913, 0.4619,
0.4157, -0.0975, -0.4904, -0.2778, 0.2778, 0.4904, 0.0975, -0.4157,
0.3536, -0.3536, -0.3536, 0.3536, 0.3536, -0.3536, -0.3536, 0.3536,
0.2778, -0.4904, 0.0975, 0.4157, -0.4157, -0.0975, 0.4904, -0.2778,
0.1913, -0.4619, 0.4619, -0.1913, -0.1913, 0.4619, -0.4619, 0.1913,
0.0975, -0.2778, 0.4157, -0.4904, 0.4904, -0.4157, 0.2778, -0.0975
};
int j,k, i;
float m;
for ( j = 0; j < 8; j++) {
for ( k = 0; k < 8; k++) {
if ( k == 0)
m = (dctmtx[k][j] * 354) ;
else
m = (dctmtx[k][j] * 248) ;
i = lroundf(m);
printf("%4d ",i);
}
printf("\n");
}
}
The first coefficient in each row appears to be to a different accuracy than the remaining:
%% convftoi
125 122 115 103 88 69 47 24
125 103 47 -24 -88 -122 -115 -69
125 69 -47 -122 -88 24 115 103
125 24 -115 -69 88 103 -47 -122
125 -24 -115 69 88 -103 -47 122
125 -69 -47 122 -88 -24 115 -103
125 -103 47 24 -88 122 -115 69
125 -122 115 -103 88 -69 47 -24
After a little finessing to find scaling factors that did give a match.
addendum
After LutzL's answer I derived the float coefficient matrix algorithmically:
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#define PI 3.14159265359
int main (int argc, char *argv[])
{
float calcmtx[8][8];
int j,k, i;
float m;
printf("float coefficients calculated\n");
for ( j = 0; j < 8; j++) {
for ( k = 0; k < 8; k++) {
if ( j == 0)
m = cos(PI*j*(2*k+1)/16)/(sqrt(2)*2) ;
else
m = cos(PI*j*(2*k+1)/16)/2 ;
calcmtx[k][j] = floorf(m*10000 + 0.5)/10000;
}
}
for ( j = 0; j < 8; j++) {
for ( k = 0; k < 8; k++) {
printf("% 2.4f ", calcmtx[k][j]);
}
printf("\n");
}
printf("\n") ;
printf("integer coefficients derived\n");
for ( j = 0; j < 8; j++) {
for ( k = 0; k < 8; k++) {
if (k == 0)
m = sqrt(2);
else
m = 1;
i = (int) (calcmtx[j][k] * 250 * m);
printf("%4d ", i);
}
printf("\n");
}
printf("\n") ;
printf("approximated integer coefficients\n");
for ( j = 0; j < 8; j++) {
for ( k = 0; k < 8; k++) {
if ( k == 0)
m = calcmtx[j][k] * 354 ;
else
m = calcmtx[j][k] * 248 ;
i = lroundf(m);
printf("%4d ", i);
}
printf("\n");
}
}
And we see that the integer matrix first coefficient is multiplied by the square root of two:
%% gencoeffi
float coefficients calculated
0.3536 0.3536 0.3536 0.3536 0.3536 0.3536 0.3536 0.3536
0.4904 0.4157 0.2778 0.0975 -0.0975 -0.2778 -0.4157 -0.4904
0.4619 0.1913 -0.1913 -0.4619 -0.4619 -0.1913 0.1913 0.4619
0.4157 -0.0975 -0.4904 -0.2778 0.2778 0.4904 0.0975 -0.4157
0.3536 -0.3536 -0.3536 0.3536 0.3536 -0.3536 -0.3536 0.3536
0.2778 -0.4904 0.0975 0.4157 -0.4157 -0.0975 0.4904 -0.2778
0.1913 -0.4619 0.4619 -0.1913 -0.1913 0.4619 -0.4619 0.1913
0.0975 -0.2778 0.4157 -0.4904 0.4904 -0.4157 0.2778 -0.0975
integer coefficients derived
125 122 115 103 88 69 47 24
125 103 47 -24 -88 -122 -115 -69
125 69 -47 -122 -88 24 115 103
125 24 -115 -69 88 103 -47 -122
125 -24 -115 69 88 -103 -47 122
125 -69 -47 122 -88 -24 115 -103
125 -103 47 24 -88 122 -115 69
125 -122 115 -103 88 -69 47 -24
approximated integer coefficients
125 122 115 103 88 69 47 24
125 103 47 -24 -88 -122 -115 -69
125 69 -47 -122 -88 24 115 103
125 24 -115 -69 88 103 -47 -122
125 -24 -115 69 88 -103 -47 122
125 -69 -47 122 -88 -24 115 -103
125 -103 47 24 -88 122 -115 69
125 -122 115 -103 88 -69 47 -24
Which matches the approximation when the float accuracy is limited.
来源：https://stackoverflow.com/questions/23843140/convert-float-to-integer