Scott, I'd rather not speculate, so could you post the code and the image you're working with?

Scott, the jpg you attached seems to have dimensions 1844x3069, which is the size you report for the post DeleteBorderComponents image. Maybe you posted the wrong starting image? Or maybe you just miscalculated the size of the starting image?

Oh, I completely misunderstood.

Hmm, I tried several manipulations of the original image before attempting DeleteBorderComponents (changing the color space, ColorNegate, etc), and every time the result is a completely black image. DeleteBorderComponents doesn't seem to have any options for setting the threshold for what's considered background, which many other image manipulation functions do, so that's kind of weird. This is the first time I've ever tried using DeleteBorderComponents, so I'd have to do some further experimentation before I can speculate on a cause for this behavior with this image. I hope someone else in the community has suggestions. Sorry that I couldn't be more helpful.

Nice explanation, Matthew!

Hi all, here comes just a little remark. I am aware that I am not actually answering the original question, but I am just assuming that "cleaning" such an image is the ultimate goal here (thank you for this interesting and challenging task!). So this is what I tried:

A sample of the original image:

I end up with:

I started by partitioning the image, working on the parts. One of the resulting advantages is that the successive processes can then be parallelized. Basically I found NonlocalMeansFilter most helpful (but it takes hours to run!). My code:

ClearAll["Global`*"]
img0 = ColorConvert["<your image>", "Grayscale"];
partImgs0 = ImagePartition[img0, {200}];
partImgs1 = ParallelMap[ColorToneMapping@*ImageAdjust, partImgs0, {2}];
partImgs2 = ParallelMap[NonlocalMeansFilter[#, 6] &, partImgs1, {2}];
partImgs3 = ParallelMap[ImageAdjust[#, {2, -.2, 1.5}] &, partImgs2, {2}];
img1 = ImageAssemble[partImgs3]

(* dirty trick: drawing black vertical lines next to left/right borders
to prevent 'DeleteBorderComponents' from swallowing whole words *)
img1data = ImageData[img1];
Table[(img1data[[n, 10]] = 0; img1data[[n, -10]] = 0), {n, 1, Length[img1data]}];
img1a = Image[img1data]
img2 = DeleteBorderComponents[img1a]
img3 = ColorNegate@DeleteSmallComponents[img2, 100]

I am sure there still is room for lots of improvement!

Regards -- Henrik

Hi Scott! Thanks for your nice reply! I hope that my use of @* was not the only reason you made the upgrade. You are right: @* means "composition", but this is not essential here; f@*g is equivalent to f[g[#]]&. But upgrading from v8 to v10 makes a lot of sens anyway! One major innovation is the introduction of Association, which is an improvement when working with large amounts of data (" There is a tsunami of handwritten material being digitized").

Regards -- Henrik

Hi Scott,

yes, I observed as well that some tiles of the partition may look "different", but typically this occurs only when tiles do (nearly) not contain any "black inc", so I was not very concerned about. My motivation was to enhance the readability of the handwriting, not to make the document as such look "nicer". My (maybe naive) idea was that the dynamic range of the image can be stretched much more by working on small partitions instead of the whole image. I would like to show a simple example for what I mean: Let's generate a test image:

ClearAll["Global`*"]
gcd = ColorData["GrayTones", "Image"];
lne = Graphics[{Thickness[.05], Black, Line[{{0, .5}, {1, .5}}]}];
img = Image @ Graphics[List @@@ {gcd, lne}, PlotRange -> {{0, 1}, {0.4, .6}}, ImageSize -> 600];
imgpart = First@ImagePartition[#, {150, 120}] &@img;

Then adjusting the whole image looks like:

the same done on tiles gives:

The fact that with ImagePartition (i.e. working on independent tiles) one can use parallelization is just a nice side effect.

Regards -- Henrik