Fix Low-Confidence OCR Words

About Fix Low-Confidence OCR Words

An OCR text layer is mostly right and occasionally wrong, and the trouble is not knowing where. This tool runs Tesseract over the whole document, then collects only the words the engine itself was unsure about (confidence below 80% by default) and shows each one next to a crop of the original image.

Corrected words are written back at the same position as invisible text with confidence set to 100; untouched words keep the original result. How OCR works and which scan settings matter are covered in the guide "Making Scanned PDFs Searchable".

What confidence catches, and what it misses

Confidence measures how sure the engine is about the shapes it saw, not whether the output is correct. Smudged ink or tiny print gets a low score and lands on the list, but characters that look nearly identical, such as the digit 0 and the letter O, or 1 and l, can be misread with high confidence. For invoices, account numbers or dates where one character matters, raise the threshold to around 90% to widen the review.

Lowering it to 60% leaves only clearly broken words and a short list. On long documents, a quick pass at a low threshold followed by a stricter pass on the pages that matter saves time.

What to fix first

A text layer exists for search and copy, so fix the words someone will search for later: names, company names, amounts, dates and document numbers. A misread "the" or "and" hardly affects search.

In tables and receipts, a column rule next to a number is sometimes read as 1, l or |, turning 48.00 into 148.00. Checking every suspect amount against its crop catches exactly this kind of error.

Worked examples

How to use it

1. Open the PDF, set the threshold and run the suspect search.
2. Compare each crop with its recognized text and fix only the wrong ones.
3. Save to get a PDF with the corrected text layer.

Frequently asked questions