Re: [Boost-users] Does anyone know how to detect image offset?

Subject: Re: [Boost-users] Does anyone know how to detect image offset?
From: Zachary Turner (divisortheory_at_[hidden])
Date: 2009-06-08 02:20:57

On Sun, Jun 7, 2009 at 9:56 PM, blp330 <blp330_at_[hidden]> wrote:

>
>
> Jens Weller wrote:
> >
> >
> > Trim the first lines and rows from both images, which only contain pixels
> > of white/transparent color. But thats surley expensive, as you have to
> > look at a lot of pixels.
> >
> >
>
> The performance issue is critical for my application. And image might
> offset
> to left, right, down. I need a way to detect the image moving.
>
> Thanks for your reply, do you have any suggestion for that?
>

How about finding the first non-white pixel in each image?  In the
"baseline" picture that you're comparing against you can pre-compute this
one time at the beginning.  Suppose you find that for the reference image
the first non-white pixel is at coordinates (r, s)

To detect a move, calculate it for the image you're trying to verify and you
find that the first non-white pixel is at coordinates (x, y).

1) It is the original image offset by some amount.  In this case, it has
been offset by exactly (x-r, y-s).
2) It is a totally different image than the original.  If you can tolerate
less than 100% confidence that the image is the same thing but just shifted,
just do a memcmp of the first few lines starting from this point.  You know
to offset into the image by (x-r,y-s) when doing the memcmps so that the
comparisons are based at the same reference.  This should give you good
confidence whether or not they are the same.

If the image is stored in a large 2-dimensional array as I suspect it may be
with only pixel color information and no other information embedded into the
cells, then once you find the first non-white pixel in each cell, just do a
memcmp against the entire rest of the buffer.  As long as the rectangles are
not different sizes, this should still work.

For example, image 1 and image 2 are both 1280x1024.  They are the same
except image1 is in the rectangle {(3, 7), (12, 18)} and image2 is in the
rectangle {(4, 8), (13,19)}.

You find that the first non-white pixel is (3,7) in image 1 and (4,8) in
image 2.  The memory from (3,7) to (4,7) in image 1 (1024 pixels) should be
the same as the memory from (4,8) to (5,8) in image 2.  It's just that the
margin has moved from one side to the other.

It realy depends a lot on the requirements of your app how to make it
faster, there are a bunch of possibilities depending on error tolerance,
needs, etc.