Many image compression standards use the DCT transform to convert spacial data into frequency data.

The DCT coefficents are then quantized based on the estimated importance of a certain frequency to the human eye.

However one thing I've always wondered about is why aren't quantization grids symmetrical across the diagonal axis? For example this is the jpeg grid at a quality of 50

enter image description here

I would think if you draw a line from the top left hand corner, to the bottom right everything above and below that line should be the same. As these coefficients are representing the same type of frequency data, but just in a different direction.

Is there something about the human visual system that causes us to see horizontal data more than vertical? As it seems that's what these grids are built for.


I understand this is a very technical question, so here's essentially the same question in simpler terms.

In image compression we try to remove any parts of the image that aren't important to how we perceive it. The charts above show the math behind it, but the logic the people who made jpeg came up with is basically this.

A horizontal pattern is more important than a vertical one. So looking at the image below you can see that jpeg says if part of the image (an 8x8 pixel block) that we want to compress looks like the one on the left, then it's very important and we need to save it.

However if the image looks more like the one on the right, it's less important and we can compress it more aggressively.

What I don't understand is they are the same pattern just rotated. Are there any studies or theories about horizontal patterns being more important to how we perceive an images quality?

enter image description here

  • Just a complete guess, but I would think maybe it has to do with the sensitivity of the human eye to vertical details not being equivalent to that for horizontal details. – twalberg Jul 30 '18 at 18:59
  • What is the resolution of your grid? – Michael C Jul 30 '18 at 21:49
  • @MichaelClark I'm using the jpeg standard of an 8x8 grid. But really I see the same thing applied to mp4's 16x16 grids as well. – YAHsaves Jul 30 '18 at 21:54
  • 1
  • 4
    I'm voting to close this question as off-topic because it already has an existing answer at another SE community that is probably a better fit for this question. Why are JPEG quantization matrices asymmetric? – Michael C Aug 6 '18 at 5:01