how do i make the imcrop function accurate to the pixel?

hi,

there is a tip in the help browser imcrop section which says:

"Because rect is specified in terms of spatial coordinates, the width and height elements of rect do not always correspond exactly with the size of the output image. For example, suppose rect is [20 20 40 30], using the default spatial coordinate system. The upper-left corner of the specified rectangle is the center of the pixel (20,20) and the lower-right corner is the center of the pixel (50,60). The resulting output image is 31-by-41, not 30-by-40, because the output image includes all pixels in the input image that are completely or partially enclosed by the rectangle."

is there a way to overcome this accuracy issue and get pixel precise output?

thanks a lot,

Answer 1

[Image Analyst]

Subtract 1 from the width and height. I agree that "center-to-center" pixel measurements can be confusing when you would really like "whole" pixels. For example, how long is this stretch of 1 pixels:

0 0 0 0 0
0 1 1 1 0
0 0 0 0 0

Is it 3 long, or 2 long? It depends on if you're considering whole pixels or going from center to center, respectively.

And they aren't always consistent on how they do it (just look at bwarea vs regionprops). I'll bring it up with them. (I've already discussed with them the annoyances we sometimes face with functions dealing with 0-1 instead of gray levels.)

[mat]

hi

any luck with mathworks?

tx mat

[Image Analyst]

No. They don't always act on my suggestions, and when they do they don't let me know in advance and I find out afterwards, like when they review my list and say "Oh, we implemented that 2 versions ago." But is my answer good enough for you to mark it as "Accepted"?

[DGM]

For what it's worth, it's easier to accept that it's unintuitive if you consider that imcrop() is not limited to working in image (pixel) coordinates. It's set up to perform the selection in an arbitrary coordinate space.

To use a non-default spatial coordinate system for the target image,

precede the other input arguments with two 2-element vectors specifying

the XData and YData:

[...] = imcrop(X,Y,...)

Also from the synopsis:

Because RECT is specified in terms of spatial coordinates, the WIDTH

and HEIGHT of RECT do not always correspond exactly with the size of

the output image. For example, suppose RECT is [20 20 40 30], using the

default spatial coordinate system. The upper left corner of the

specified rectangle is the center of the pixel (20,20) and the lower

right corner is the center of the pixel (50,60). The resulting output

image is 31-by-41, not 30-by-40, because the output image includes all

pixels in the input that are completely or partially enclosed by the

rectangle.

I've thought before about writing a size predictor or parameter selection aid for imcrop(), but I don't know who would actually need it. If you want to select an image region in pixel coordinates with utter certainty that you get what you think you asked for, just use direct array indexing instead of imcrop(). At least that way you can have the opportunity to properly handle edge collisions without just getting an empty array in return.