I wrote a program to clean up scans of handwritten notes while simultaneously reducing file size.

Example input and output:

Left: input scan @ 300 DPI, 7.2MB PNG / 790KB JPG. Right: output @ same resolution, 121KB PNG.1

Disclaimer: the process described here is more or less what the Office Lens app does already, and there’s probably any number of other tools that do similar things. I’m not claiming to have come up with a radical new invention – just my own implementation of a useful tool.

If you’re in a hurry, just check out the github repo, or jump down to the results section, where you can play with interactive 3D diagrams of color clusters.

Some of my classes don’t have an assigned textbook. For these, I like to appoint weekly “student scribes” to share their lecture notes with the rest of the class, so that there’s some kind written resource for students to double-check their understanding of the material. The notes get posted to a course website as PDFs.

At school we have a “smart” copier capable of scanning to PDF, but the documents it produces are… less than attractive. Here’s some example output from a handwritten homework page:

Seemingly at random, the copier chooses whether to binarize each mark (like the x’s), or turn them into abysmally blocky JPGs (like the square root symbols). Needless to say, we can do better.

We start out with a scan of a lovely page of student notes like this one:

The original PNG image scanned at 300 DPI is about 7.2MB; the same image converted to a JPG at quality level 85 is about 790KB.2 Since PDFs of scans are typically just a container format around PNG or JPG, we certainly don’t expect to reduce the required storage size when converting to PDF. 800KB per page is pretty hefty – for the sake of loading times, I’d love to see things closer to 100KB/page.3

Although this student is a very neat note-taker, the scan shown above looks a bit messy (through no fault of her own). There’s lots of bleed-through from the opposite side of the page, which is both distracting for the viewer and hard for a JPG or PNG encoder to compress, compared to a constant-color background.

This is what the output of my noteshrink.py program looks like:

It’s a comparatively tiny PNG file, weighing in at just 121KB. My favorite part? Not only did the image get smaller, it also got clearer!

Here are the steps required to produce the compact, clean image above:

Identify the background color of the original scanned image.

Isolate the foreground by thresholding on difference from background color.