How to find out what part of image is less lens-distorted?

Question

I'm developing an application which uses a Canon DSLR to sequentially shoot an object from different views. Often the object is too big to fit in frame so I have to make more than one series of photos.

It's really important for the photos to be distorted as little as possible (the photos are then merged into one and result image is used in mathematical algorithm). For that purpose it's really important to know which part of image is least lens-distorted.

I suppose the least lens-distorted rectangular part of a shot is defined by focal length of the lens I'm using, but I'd like to know which part it is exactly, or at least how to calculate that part.

Could anyone help me in clarifying this?

EDIT:

The object always has a shape of cylinder. It is placed between two shafts which rotate the object (rotation is controlled by software). So, the goal is to get an image that represents the surface of this object.

The process is as follows:

1. I set the number of rotations to be made - call it 'n'. On each iteration the object is rotated by 360/n degrees, and for each iteration camera shot is made.
2. Once n rotations (and shots) are made, if object doesn't fit in camera frame, carriage, that holds tha camera shifts to the side and repeats this process for the new portion of the object. Otherwise process finishes.

Here's an example of photo for better understanding:

Algorithm takes a part of image, which represents 360/n degree part of object's surface and saves it. After finishing the process, those pieces are merged into one and the result looks something like that:

I hope this clarifies the process a bit more.

Answer 1

I suppose the least lens-distorted rectangular part of a shot is defined by the focal length of lens I'm using, but I'd like to know which part it is exactly, or at least how to calculate that part.

I think you really need to start here:

What is the difference between perspective distortion and barrel or pincushion distortion?

... because there are several different types of distortion and you really need to be clear before you can do anything about it.

The basic answer is: perspective distortion entirely depends on the distance you are from the object, and therefore focal length is a factor only so much as it affects your framing. For this, there is no clear line between "close up off-center things look weird" and something which our brain sees as normal.

For distortion that is actually a lens defect (pronounced "design compromise"), there is no general answer because you'll need to characterize each lens (or find an existing "lens profile" which includes this characterization). For zoom lenses, this will likely change as you change focal length, but this is due to the design of each particular lens's zoom mechanism, not inherent to focal length.

Camera lenses sold for photography are generally designed for uses that are more art than science. The design intent is to produce an appealing rendering. Therefore, they are not always ideal when pressed into service as measuring instruments.

Back to the beginning: if you need to move the camera to fit more in the frame, that inherently means you are changing perspective. If your subject is not flat, well, there's nothing to be done about this. If you cut out the parts which are distorted relative to your first shot, you'll cut out the information you gained from moving. Your software can twist things after the fact to align, but can't make up for this.

(The only solution is to take many pictures from slightly differing views, build a 3D model from that, and then render the model rather than working with the raw data from the lens. That's what our brains do, after all.)

Answer 2

It seems that the camera will be in a (fairly) fixed position, while the item you're photographing will always be the surface of a cylinder. The "distortion" in this case is not caused by the lens, but by the shape of the cylinder.

The way scanners handle this problem is to scan a single pixel strip that runs the length of the cylinder. This would require that the motor be capable of rotating the cylinder in increments that match the width of a single pixel. If it is not able to do that, you can use a wider portion of the image.

Then as long as the strips don't overlap, you can just append them to create the final image. For side-to-side camera movements, just repeat the process and append the images appropriately.

If you can control the camera and drum movements so that only a single pixel from each photo is used, you'd be mimicking how a drum scanner works. A single pixel capture from the center of the frame, when using an appropriately aligned lens, would be expected to have the least distortion possible from any particular lens-sensor combination.

Answer 3

It is not clear what you call "lens-distorted". A perfectly accurate rectilinear lens will distort angles and proportions as you move from the center of the image: that's a property of the projection rather than the lens, and obviously there is no way around taking the projection into account for your application.

Modern cameras and lenses communicate the deviations of the lens from this ideal even with interchangeable lenses (but certainly so with cameras having fixed lenses) and the camera compensates for it. To check whether it does, take some photographs using the widest setting of the lens and compare the uncorrected(!) raw image (this may require switching off geometric correction in your raw image processing software) with the JPEG.

More expensive lenses, especially with a moderate zoom range or even a fixed focal length, tend to require very little additional correction but either way you should be able to rely on the results to be rectilinear to a sensible degree (unless you use an intentionally non-rectilinear lens like a fisheye, of course).