Understanding gain of the DCT transform

Question

I have been reading the guide from:

• JPEG Huffman Coding Tutorial

I eventually reached the point where it states:

Finally, we want to convert the DCT DC value to an RGB value. Assuming that the gain of the DCT transform is 4, we divide the values by 4 to get block1 = -128, block2 = +127.

I have to admit, I do not understand where this magic value '4' is coming from. With the help of djpeg and jpegtran, I was able to craft two different JPEG files:

1. white512.jpg () is a JPEG where the DC value is +512,
2. white338.jpg () is a JPEG where the DC value is +338

Using a simple check, we can confirm they both decompress to the same white ppm file:

% djpeg -outfile 338.ppm white338.jpg
% djpeg -outfile 512.ppm white512.jpg 
% md5sum 338.ppm 512.ppm
18f695bae266c6a24506936dab1dfc23  338.ppm
18f695bae266c6a24506936dab1dfc23  512.ppm

Could someone please tell me:

1. Where is this value '4' value coming from ?
2. How come both the DC value 338 & 512 expand to the same white value (0xFF) ?

---

For reference, the JPEG files uploaded on imgur have been edited in my back. In particular the upload process strips the APP0 marker. To get the exact same image as I have on disk, one can apply back the default jpegtran settings:

% jpegtran -outfile white512.jpg -optimize jTKco.jpg
% jpegtran -outfile white338.jpg -optimize WqThz.jpg
% md5sum white512.jpg white338.jpg
695527adb6b269214e38e45157edb256  white512.jpg
051434f11491bdf1c1be38ad8f69f5be  white338.jpg

Answer 1

Here is the small experiment I was able to conduct on my Debian system. I crafted two small $8 \times 8$ JPEG files again:

• white: ()
• black: ()

I used the following commands:

% convert -size 8x8 -depth 8 xc:white white.pgm
% convert -size 8x8 -depth 8 xc:black black.pgm

then:

% cjpeg -sample 1x1 -quality 100 -grayscale -outfile white.jpg white.pgm
% cjpeg -sample 1x1 -quality 100 -grayscale -outfile black.jpg black.pgm

Note that in this case, it is a single component JPEG (no color space transform):

% file white.jpg
white.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1, segment length 16, baseline, precision 8, 8x8, components 1

Upon inspection, one can check that the DC coefficient stored in white.jpg is 1016, while the DC coefficient stored in black.jpg is -1024. Which can also be represented as:

$$\begin{bmatrix} 1016&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0 \end{bmatrix}$$

One can also check that the quantization table is simply a a bunch of ones:

% djpeg -verbose -verbose -outfile dummy.ppm white.jpg
[...]
Define Quantization Table 0  precision 0
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1
           1    1    1    1    1    1    1    1

So taking the entry-for-entry product with the quantization matrix from above results simply in the same matrix.

Going back to the equations on wikipedia page:

$$f_{x,y} = \frac{1}{4} \sum_{u=0}^7 \sum_{v=0}^7 \alpha(u) \alpha(v) F_{u,v} \cos \left[\frac{(2x+1)u\pi}{16} \right] \cos \left[\frac{(2y+1)v\pi}{16} \right]$$

Simply gives on our case (no rounding error):

$$f_{x,y} = \frac{1}{4}\left(\frac{1}{\sqrt{2}\sqrt{2}}1016\times 1+63\times 0\right) = 127$$

So adding 128 to each entries lead to:

$$\left[ \begin{array}{rrrrrrrr} 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 & 255 & 255 & 255 & 255 \end{array} \right]$$

Repeating the above for the black.jpg image with DC value -1024 gives:

$$f_{x,y} = \frac{1}{4}\left(\frac{1}{\sqrt{2}\sqrt{2}}\times -1024\times 1+63\times 0\right) = -128$$

Which is what we expected.

Answer 2

Multiplied by the first value of the Quantization Table.

jpeg_component_info->quant_table->quantval[0]