Why resolution decides whether OCR works
Most OCR is run on whatever render the page viewer already had on screen — an image sized for a monitor, somewhere around 96 DPI. The engine never sees the document, only that picture of it, and by then the detail it needed has already been thrown away. Everything downstream inherits the loss.
Work the arithmetic. Ten-point type stands about a seventh of an inch tall, so at 96 DPI each character is given roughly thirteen pixels of height, and thirteen pixels is not enough to hold apart rn and m, or 0 and O. At 300 DPI that same character has about forty-two pixels to describe itself.
So the page is rendered at 300 DPI before recognition begins, rather than at the size you happened to be viewing it. That one step is usually the difference between output you correct and output you retype from scratch. It cannot recover detail that was never captured — a photograph taken out of focus stays out of focus.
What the cleanup step actually repairs
Raw recognition hands back the characters it saw, line by line, and that is not the same thing as text you can use. A word split across a line end arrives as two fragments, so a search for evaluation finds nothing. Tables collapse into one run of numbers with the columns gone.
The cleanup pass then reads that output as language rather than as shapes. Hyphenated breaks are rejoined, rn and m are settled by the words on either side, and a 0 against an O is decided by its neighbours — digits in a GST invoice column, letters in the supplier's name above it.
It repairs what is recoverable, which is a real limit. Where the scan holds enough of a character, context settles the reading; where it holds nothing, there is nothing to reason from and the gap stays a gap. Multi-column layouts and merged table cells still need your eye on the ordering before you trust it.
Why the per-page confidence score matters
The usual output is one continuous block with nothing marking which parts were read cleanly and which were guessed. Forty pages of that leaves two options, both poor. Read every line back against the original, or trust the lot and find out in a fortnight that ₹4,80,000 came through as ₹4,80,00.
Each page here carries its own score instead. Four weak pages out of forty tells you where the next twenty minutes should go, and those four are usually the ones you already suspected — the faxed annexure, the page photographed at an angle, the third-generation photocopy of a land record.
Treat a low score as an instruction and a high one as the absence of an instruction. A genuinely illegible scan can still produce text that looks entirely ordinary — correctly spelled, grammatical and wrong — because plausible words are what a cleanup step is built to produce. Check the low pages against the paper.