I dunno if this should go in a Math forum or a Programming forums, but I'll post it in both and see where I get.

I have two computer images... one of them is an "original" image (a large TIF file). The other one is a transformed version of the original image... it's been rotated, sheared and translated in a software program. I need to do some work on the transformed image, but I need the (x-y) coordinates of each pixel in the original image to finish my calculations.

I know that the image was rotated and sheared with a 3x3 Transformation matrix. If I had the matrix, I could derive the second image from the first (or vice-versa) myself. I don't know exactly how much it was rotated, sheared, or translated, so I can't just derive the matrices from a set of known transformations. What I do have is a set of corresponding points (the corners, et al) in each image, and their corresponding (x,y) coordinates. So here's my dilemma:

Using a set of corresponding transformed points ((x,y) -> (x',y'), three or more of them), can I derive the Transformation matrix that was u开发者_JS百科sed to turn one image into the other? If I can derive the matrix, I can solve for the original coordinates of all the pixels (all 18-million of 'em) and get the calculations done that I need to do.

Can anyone help? I'm familiar with linear algebra... just not familiar enough to derive this without a whole lotta head scratching. Anything is appreciated!

Mike

Not sure if you want manual or automatic.

Manual

If you specify the transformed coordinates of the four corners of your rectangle, then you can derive the transformation equations:

x' = c1 * x + c2 * y + c3 * x * y + c4
y' = c5 * x + c6 * y + c7 * x * y + c8

(From Pierre Wellner's Interacting with Paper on the DigitalDesk, page 67)

Now you just have to solve for the coefficients of the equation.

With four point pairs, the two sets of four simultaneous linear equations can be quickly solved by Gaussian Elimination to find the values of c1-8.

Lastly, you can turn those equations into the 3x3 matrices you want. The above equations are powerful enough to do non-linear transformations and you can simplify it into the 3x3 affine shear matrix.

But I would just stick with the nonliner equations (above) since they can handle perspective distortion.

Automatic

Same method, but you can use an edge-detector comboined with a line detection algorithm to find a set of 4-ish lines that makeup a rectangle.

If your image rectangles really stand out (whiteish images on a dark background), then you can use corner detection available from libraries like OpenCV's Feature Detection (see cv::cornerHarris).

You can intersect those lines to find the four corners and use the transformation equation.

I think you should start by providing a list of, say 3 points (for 6 unknowns) with X/Y coordinates before and after transformation.

Then somebody more clever than I should pop that into a set of linear equations and then feed it to (say) Wolfram Alpha for solving.

The top of Java's documentation for AffineTransform shows how the matrix needs to be set up:

[ x']   [  m00  m01  m02  ] [ x ]   [ m00 x + m01 y + m02 ]
[ y'] = [  m10  m11  m12  ] [ y ] = [ m10 x + m11 y + m12 ]
[ 1 ]   [   0    0    1   ] [ 1 ]   [           1         ]

Removing most of the fluff leaves:

[ x']   [ m00 x + m01 y + m02 ]
[ y'] = [ m10 x + m11 y + m12 ]

Then you just set up a set of 6 x 2 equations like this:

m00 x + m01 y + m02 - x' = 0
m10 x + m11 y + m12 - y' = 0

(repeat for 2 other x/y before/after pairs)

and throw them at an equation solver.

You only need 3 points to define a 3x3 transformation matrix. If you have the points (0,0), (0,1) and (1,0) and transform them by the matrix [a b c d e f 0 0 1], you'll get (c,f), (b,e) and (a,d).