2026-09-02
The engine slices the image into small regions and judges which character each most resembles. Its output is probabilistic — the cleaner and more regular the input, the better the result.
1. Scan at 300 DPI or higher. 2. Deskew first — even two or three degrees hurts badly. 3. Enable image enhancement (grayscale, contrast, binarization). 4. Pick the right language pack — for mixed Chinese/English, use the combined model.
5. Avoid patterned paper and watermarks. 6. Crop tables into regions before recognizing. 7. Use the original scan, never a photo of a screen.
Try selecting text in the PDF first: if you can select and copy, the file already has a text layer — copy it or convert to Word directly. Running OCR anyway only introduces new recognition errors. OCR is truly needed for image-only scans.
No engine reaches 100%. When proofreading, prioritize numbers, amounts, names and punctuation — they are both the most error-prone and the most consequential.