Close
How do you stitch scans?
I usually first try autostitch:
http://www.cs.ubc.ca/~mbrown/autostitch/autostitch.html
It's fully automatic, but only support jpg in/out and cannot lock lens paramenters to flatbed scanner (equivalent to a camera lens shooting from infinite far away from the scan). So sometimes it gives me erratic results.

So if it doesn't work I'll try hugin
http://hugin.sourceforge.net/
It can set focal length manually but is not as automatic.

So what's your way?
I use photomerge in PS to begin with..of course Photomerge isn't perfect..so i fix the stitch manually by hand with transform and some editing (usually gaps,folds). CS3 has still the best algorythm for stitching imo.
Microsoft ICE sometimes can do things PS can't do
I always do my manually in Adobe Photoshop CS2. I've played with some of those auto stitching programs before, but I like doing the manual work to stitch my scans. Cause from there, I can apply any kind of texture or gradient i need along with color alterations from the mistakes of the scanning process.

My tool in CS2 comprise of these:

  • The brush tools
  • The Smudge tool
  • The Clone Tool
  • The Pen tool
  • The History Brush (Set your history to right after you just joined the 2 images)
You can always use some of the special features on CS2 to help you, but to the naked eye, most users won't see the little details anyway.
The simplest method.
Repaint around seams with brush/airbrush, that's all =D
I always have difficulty stitching manually, since small rotation error can cause lines to mismatch which is impossible to cover up by retouching tools...
So how do you cope with this? find the correct rotation by trying or PS can find it automatically for you?

PS: thanks for telling me about MS ICE and PS photomerge, i'll play around them later
Just free transform to line up. Put each part in a separate layer, and find a spot near the top of the image that lines up; if there's no overlap, it's easiest with two straight lines that cross the boundary in different directions, so you can line it up on both axes, but sometimes it takes some guessing.

Then free transform on one of the layers, drag the anchor point (the dot that appears in the center of the layer) to the point near the edge that's most correct, scroll down to the bottom edge which is off due to the rotation, and change the rotation (don't drag the transform outline; select the rotation input at the top and press up/down to change .1 degree at a time) to line it up. Since the anchor is at the spot that's lined up, it won't break that side of the image.

If both parts are unwarped, this usually can get it very close. If you only have one line crossing across, it can be trickier: you can line up the line, but not know how big the gap should be, and it can take some experimentation. Sometimes you only have rounded eges to work with, which is trickier.

One major limitation in Photoshop is that you can only rotate in free transform by 0.1 degree; sometimes this is too coarse, so 0.3 will be rotated too little and 0.4 too much; 0.35 will just round to 0.4. This is really annoying. If you use Image/Rotate Canvas/Arbitrary, you can rotate by much finer increments, but that will only rotate the entire image and not one layer, so you have to split the layers apart, and you lose the benefit of being able to see the preview in realtime.

If your image has strict horizontal or vertical lines, it's a big help. Keep each part in a separate image, select the ruler tool, drag it along the sharp edge. Zoom in and line it up exactly, then hit rotate/arbitrary and it'll fill in the rotation needed to make the ruler horizontal or vertical. Repeat for the other part. If both images have an edge like this, then it'll rotate both exactly, so you only have to line them up. (post #57830) Beware of doing this with the edge of the paper rather than a feature in the actual printing (the print itself may be slightly off), but it's also worth trying if you're having trouble.
Like Peto said. Most of the scans i've worked on are usually pretty close in alignment so all i have to really do is work on the mid section of the 2 scans.

kiowa - I do not believe PS (well CS2) has an ability to do this. I do believe most of them has to come from the manual use of the Free Transformation tool Peto talked about.
petopeto said:
One major limitation in Photoshop is that you can only rotate in free transform by 0.1 degree; sometimes this is too coarse, so 0.3 will be rotated too little and 0.4 too much; 0.35 will just round to 0.4. This is really annoying. If you use Image/Rotate Canvas/Arbitrary, you can rotate by much finer increments, but that will only rotate the entire image and not one layer, so you have to split the layers apart, and you lose the benefit of being able to see the preview in realtime.
Use Gimp 2.6 =D
free degree rotate, pan out side of canvas, high quality preview anytime, any ratio. e.t.c.
I both have PS & Gimp, but i always use Gimp to aligh images ;)
You can pan outside the canvas in PS by switching to fullscreen (press F). I always (100%) work in fullscreen mode.
Another trick, if you don't have anything flat to align and you're having trouble figuring out the correct rotation: align the screening.

The screening, for any particular ink color (CMYK), is at a particular angle. This is probably the same on both pages. Use the ruler tool, and align it against a particular line of dots. Try to pick an area where a single color is clearly visible (different colors are often aligned to different angles), and where you have a long, clear shot--the longer the ruler is, the less error. Note the angle in the info pane. Do the same thing to the same color in the other image, then subtract them, and rotate one of them by the result (or the inverse, depending which way you're going).

The info pane, unfortunately, only shows one tenth of a degree, when an exact alignment could really use a hundredth, but it can get you pretty close. Once you have the correct angle, also getting the position aligned often becomes much easier. This only works on unfiltered raws, of course, and it won't work on high-resolution FM screened printing.

(Actually--more advanced: you can get a number to the hundredth by opening rotate arbitrary; it'll be modified, since it's trying to snap to an axis, but the same subtraction will still probably work. But you'll need to yank the parts back into separate images and use rotate arbitrary to apply a rotation with that precision, since Free Transform seems to round to .1 degree.)
Well, just reread all the replies, thanks to eyerybody.

Just spent another few of hours on automated tools.

Autostitch is great, but it seems that I tend to get large barrel error, and has very high probability of running into the "out of memory" error.

Also it lacks lossless output, which is fatal for obsessive-compulsive perfectionist syndrome victims. (and in fact, saving unfiltered raws into jpeg sometimes do give you problems)

Microsoft ICE is also quite well, but sometimes I get jagged seams. It's ideal if it works, but if it doesn't, you can't do anything with it.

Hugin is highly customizable and nearly guaranteed to work-- as long as your raws are not distorted too seriously and you have enough time refining your control points. And you get numerical estimates for misalignments on different parts of your image in pixels. Compared with manual aligning and fixing, this software spreads distortion over the whole canvas instead of concentrating them to a few feature lines, which may give me some more peace of mind when stitching -_-||.

Also if your scanner's light is not quite uniform, its auto exposure levelling can save your ass.

But it's REALLY SLOW when computing output and has quite some weird behaviour-- sometimes you get stuck to some wrong output so save project often. Also, it takes quite a while to learn.

PS: according to a tutorial on using hugin to stitch flatbed scans, field of view angle can be set as you like, and the author used 40 degrees. But I think this angle should be set to a small value, like 1 deg. Because scanning is equivalent to a camera lens shooting from infinite far away from the scan.

I think I'll try manual ways tomorrow.

Ok, going to bed, it's 3am in my timezone...