How to Unblur and Read Pixelated Text in Scanned Documents

My grandmother left behind about eighty letters from the 1940s when my grandfather was overseas. Beautiful, flowing handwriting on thin airmail paper. The problem was that the letters had been scanned at some point, probably a decade ago, at what looked like 150 DPI on a cheap flatbed scanner. About a third of the words were unreadable. Individual characters were smeared, ink had faded to pale gray, and the thin paper meant the writing on the other side bled through and created a ghostly double image.

I tried reading them as is. Gave myself a headache after about three letters. Then I started looking for ways to make scanned text readable again. Here is what worked and what was a waste of time.

Why scanned text is different from photo enhancement

Most AI upscalers are trained on photographs. They are very good at faces, landscapes, and objects. Text is different. A letter is a very specific shape. If the AI rounds a sharp corner or fills in a gap incorrectly, it turns one letter into a different letter. A lowercase "c" becomes an "o". An "rn" combination becomes an "m". These are not cosmetic problems. They change the meaning of words.

The ideal approach for text is different from photos. For photos, you want the AI to generate realistic detail. For text, you want it to sharpen edges and remove noise while changing the character shapes as little as possible. It is more about cleaning up what is there than generating what might be there.

What I tried and how it went

First I tried Photoshop. The sharpen filter made the text slightly more defined but also introduced halos around every letter. After sharpening, the text looked like it had a faint glow. It was more legible but uncomfortable to read for more than a few minutes at a time.

Then I tried a dedicated photo upscaler. It made the letters look cleaner but slightly rounded their shapes. An "e" that should have had a sharp crossbar had a soft, indistinct one. Fine for casual reading, not good enough for preserving historical documents where the exact wording matters.

What finally worked was using a general AI upscaler at 2x with conservative settings. The 2x upscale doubled the pixel dimensions, which on its own made small text easier to read. The AI sharpening was subtle enough that it did not distort letter shapes. The result was not perfect. Maybe 85% of the previously unreadable words became readable. The remaining 15% were words where the original scan was so degraded that no amount of enhancement could recover them.

The OCR pipeline that saved weeks of work

After making the letters readable, I wanted to create searchable text. I ran the enhanced scans through OCR software to digitize everything. Here is what I learned about making this work well.

Always enhance before OCR. The difference in recognition accuracy is dramatic. On my grandmother's letters, OCR on the original scans got about 60% of the words correct. On the AI enhanced versions, it got about 85%. That 25% improvement meant I was correcting a few words per page instead of typing every other word from scratch.

For documents where every word matters, like legal papers or genealogical records, do not trust OCR alone even on enhanced scans. The AI and OCR together will get most of it right, but a single misread character in a name, date, or address can send you down the wrong research path. Use the OCR as a starting point and verify by reading the enhanced image yourself.

What to do about bleed through from the other side

The thin airmail paper in my grandmother's letters meant the writing on the back showed through. This confused the AI upscaler. It tried to sharpen both layers of text simultaneously, turning the ghost writing into a second set of smeared characters overlaid on the first.

The fix was to preprocess the scans before enhancement. I opened each scan in an image editor and increased the contrast until the background was as close to white as possible. This reduced the visibility of the bleed through text. Then I ran the AI upscale. The result was not perfect. The ghost text was still faintly visible. But it was subdued enough that the front side text was clearly dominant.

For the worst cases, where bleed through was so severe that both sides were equally visible, I duplicated the layer, applied a heavy blur to isolate the ghost text, and subtracted it from the original. This is getting into manual restoration territory, but for the five or six worst letters out of eighty, it was worth the effort.

I finished digitizing all eighty letters in about three weekends. TheClarifyPix AI enhancement made most of them readable. The OCR turned the readable images into searchable text. The manual work was verifying names, dates, and places. My grandmother's letters are now a searchable PDF that my whole family has copies of. The original scans were nearly unusable. The enhanced versions brought them back to life.