2002.01.21 23:26 "Re: How to interpret 16-bit GrayScale image?", by Martí Maria

Hi,

Don't trust blindly in me, just see by yourself. :-)

Here is a small program to acotate the induced error by the 256 method ( << 8 for expanding to 16 bits, >> 8 to convert to 8 bits) and the 257 method (multiply and divide by 257).

To check error, I sweep from 0 to 0xFFFF, quantizing to 8 bits and then restoring back to 16 bits. The result is next compared to original value and the error is computed as abs(quantized - original). I avoided stnadar deviation and further stadistics since seems clear we want peek error only.

Using 257 method, you got a maximum error of 128, that is almost half 8-bit value. Using << 8 you got 255, that is almost one 8 bit value.

FEFD / 101 = FD     are you sure you want this?

Yep, you have not done the neccessary rounding!:

FEFD / 101 = 253.996, rounding 254 = 0xFE

Just to make sense: 0x33FF is obviously more close to 0x34 tha 0x33.

Regards,
Mati.

------ cut -----

#include <stdio.h>
#include <stdlib.h>

int main()
{
 int i;
   unsigned char a8_by256, a8_by257;
   unsigned short a16_by256, a16_by257;
   int error256, error257;
   int max256 = 0, max257 = 0;

 for (i=0; i < 0xFFFF; i++) {

    a8_by256 = (unsigned char) (i >> 8);
      a8_by257 = (unsigned char) (double) (i / 257. + .5);

      a16_by256 = ((unsigned short) a8_by256 << 8);
      a16_by257 = ((unsigned short) a8_by257 << 8) | a8_by257;

      error256 = abs(a16_by256 - i);
      error257 = abs(a16_by257 - i);

      if (error256 > max256) max256 = error256;
    if (error257 > max257) max257 = error257;
   }

   printf("Error (Max) by 256 = %d, by 257 = %d\n", max256, max257);
   return 0;
}

----- cut----