Diagnose blank OCR, scrambled reading order, wrong language, missing lines, and factual translation errors without guessing at the source.
A blank, scrambled, or confident-looking wrong result usually has a diagnosable cause. Use these guides to isolate the image, OCR, language, and layout stages instead of repeatedly submitting the same file.
If the result is completely empty, start with what to do when OCR returns no text.
Do not diagnose a translation by reading only the final sentence. Compare the original image, prepared image, OCR text, and translated text in that order. The first record that differs from the evidence identifies the stage to fix. A wrong digit already present in OCR is not a translation error. Correct words in the wrong sequence point to layout segmentation. Accurate source text with the wrong meaning points to translation or missing context.
| Symptom | First isolation test | What the result tells you |
|---|---|---|
| No text returned | Submit one upright, tightly cropped, high-contrast line from a supported image. | A successful control points to the original image size, orientation, format, or layout. |
| Missing or substituted characters | Compare the smallest failed character at full image size and in a new close capture. | Improvement after recapture points to lost visual detail rather than language choice. |
| Correct words in the wrong order | Extract each column, bubble, or rotated region separately. | Correct regional results show that page segmentation, not word recognition, failed. |
| Fluent translation with wrong facts | Audit digits, names, units, symbols, and negative words in the OCR text. | A mismatch reveals whether the factual error entered before translation. |
| Correct text but clipped layout | Place the actual translation into a flexible copy of the design. | Overflow confirms a production constraint rather than a recognition problem. |
Keep the original file unchanged. Create a new test crop containing one clear line and note exactly what changed, such as orientation, crop, lighting, or file source. Compare the OCR outputs character by character. Changing several variables at once may produce a better result without showing which change solved the problem.
Use a known clear sample to distinguish an input problem from a temporary service problem. If the clear sample also fails, keep the source files and try again later rather than repeatedly editing the difficult document. If only the difficult page fails, follow the empty OCR diagnostic or the reading-order repair method that matches the observed symptom.
Mark a field unreadable when independent captures do not reveal its strokes, the original is unavailable, or processing creates several equally plausible characters. Do not choose the version that produces the smoothest sentence. Request a better source, consult a person who can inspect the original object, or obtain an authoritative copy. A visible uncertainty is safer and more useful than a confident guess.
Pick the first observable failure. A blank result calls for input diagnosis, while correct words in the wrong sequence require layout repair before translation begins.
An empty result does not prove that an image contains no text. It means the recognizer could not return usable text from that input and configuration.
A tight crop of one clear line is the fastest way to isolate file, scale, layout, and service problems.Correct words in the wrong sequence produce an incorrect translation. The repair belongs between recognition and translation.
Reconstruct the source reading order before translation, using labeled crops as the audit trail.