Turn Chinese Images into Editable Text

A screenshot of a Chinese product label may contain exactly the information you need, yet none of its words can be selected. Optical character recognition, or OCR, turns the visible characters into editable text. The useful result is a checked transcription: a name you can search, a paragraph you can study, or a quantity you can copy without retyping it.

Start with the clearest image, choose the language actually printed on it, and recognize one meaningful area at a time. Then compare the result with the image before translating, converting between simplified and traditional Chinese, or putting numbers into another document. A fluent-looking sentence can still contain a wrong character.

Start with the best available image

First try selecting the original text. A webpage or PDF may already contain a usable text layer even when a screenshot does not. Copy a short passage and compare it with what is visible. If the result is accurate, you can avoid a recognition step. Use OCR when the information is available only as pixels, or when the existing text layer is unusable.

Choose the original screenshot or full-size photograph rather than a thumbnail forwarded through several messages. Zoom in enough to inspect the small strokes in the Chinese characters. If those strokes have merged or disappeared, enlarging the same blurry image will not recover the missing information; obtain a clearer source where possible.

For a new photograph, place the page flat, keep the camera square to it, and use even light. Check that shadows, reflections and a curved book gutter do not cover the passage you need. A sideways image needs rotation; a page that narrows toward the top needs a better photograph or perspective correction in an image editor.

The ChinaClue tool accepts PNG, JPEG and WebP images, with limits of 15 MB and 24 million pixels. If your source is a PDF, export the required page as an image first. For an oversized image, isolate the useful passage in an editor before selecting it. Keep a small clear margin around the characters rather than cutting through their strokes.

Match the printed language and layout

The website language and the recognition language serve different purposes. You can use an English interface to recognize a traditional Chinese notice. In the OCR tool, choose Traditional Chinese + English because that describes the document, even if you later want the text in simplified Chinese.

Use Simplified Chinese + English for a simplified passage, or the combined simplified, traditional and English option for a genuinely mixed passage. Use English for an English-only image. Selecting more languages is not a substitute for a sharper image, and it may require downloading additional recognition data.

Consider a label containing 净含量 and a traditional version containing 淨含量. These mean net content, but the first character is printed differently. OCR should first recover the characters present in the source. Selecting the traditional model is not a command to convert every result into traditional Chinese, nor is selecting the English interface a request for an English translation.

Layout choices describe the arrangement, not the subject. A price label is not automatically scattered text; a tightly grouped label may work better as one block. Start with the setting that matches the selected area, and compare a second setting only when the first result has a specific problem.

Recognize one complete area at a time

Open the Chinese image-to-text tool and select the image. Inspect the preview before starting recognition. The left and right rotation controls turn it by 90 degrees. Rotate first, then select the recognition area, because rotation resets the crop. A slight diagonal tilt needs correction in an image editor rather than another quarter-turn.

Drag across the preview to select a region, or adjust the left, top, width and height percentages. Include a complete paragraph, label or column. Cropping only the number can remove the unit or heading needed to interpret it. For a two-column page, recognize the columns separately if the combined output mixes their lines.

Choose the printed language and layout, then select Recognize text. The browser loads the recognition data and processes the selected image area on your device. Model downloads still require a connection, so browser processing does not mean that the page is always available offline. Leave the tab open while recognition runs; use Cancel if you need to stop it.

The recognized text is editable. Save or copy any corrections you want to keep before running recognition again, because a new run replaces the result. When working through several regions, put each checked passage into your own document in the intended reading order. Keep a heading or page number with it so that you can find the source again.

Check meaning-bearing details first

Proofread with the image and text visible together. Start with details where a small error changes the task: names, dates, negation, quantities, units and reference codes. Then read the whole passage for missing lines and sentence boundaries. Familiar wording can help you notice a problem, but it should not be used to invent a character that the image does not clearly show.

The following label is an invented practice example, not an actual product record. Keep its line labels beside the values when checking the transcription:

名称：绿茶
净含量：1.25 kg
编号：0018
单价：38.50 元

If 1.25 kg becomes 125 kg, a lost decimal point changes the quantity by a factor of 100. The correct value is 1,250 grams, which provides a useful independent check when another document uses grams. If 0018 becomes 18, the numeric value may look equivalent, but the reference code no longer has the same four characters.

Check 38.50 character by character, including the decimal point and final zero. Confirm that 元 belongs to that amount rather than another nearby value. An OCR result may contain recognizable numbers while joining them to the wrong row. Rechecking a calculation cannot repair a transcription that paired the wrong label and value.

For Chinese text, compare similar shapes carefully: 未 and 末, or 土 and 士, differ in stroke proportions. In an alphanumeric code, compare the digit 0 with the letter O. These are proofreading examples, not a claim that every image will produce those errors. Preserve punctuation and a word such as 不, which can reverse the meaning of an instruction.

If you do not read the passage confidently, use OCR to make it searchable, then check the uncertain parts with a competent reader or the document's issuer. A translation of incorrectly recognized text can sound convincing while preserving the original recognition error. Complete the transcription check before treating a translation as an explanation of the source.

Repair the specific recognition problem

An empty result calls for checking the selected region first: it may contain mostly blank space, a photograph, or text that is too small to resolve. Return to the whole image, identify the required passage and make a new crop. Confirm the language selection and orientation before changing several settings together.

If characters look plausible but the lines are in the wrong order, treat it as a layout problem. Split a multi-column page into columns or separate a heading from its surrounding labels. For a table, recognize a manageable row or group of cells and reconstruct the table while checking against the image. Plain text output does not preserve a spreadsheet's cell relationships.

If only one word is unclear, crop that word with a little surrounding context and compare the new result with the first. Keep both readings in your working notes until the image or another reliable source resolves the difference. Repeating recognition on the identical blurry input is not independent confirmation of a disputed name or code.

When the problem is glare, heavy shadows, decorative lettering or damaged print, a better source may save more time than repeated OCR attempts. Record an unresolved passage in your working copy instead of silently guessing. For a short readable label, manually typing it and checking every character can be the more dependable route.

Keep a checked copy before conversion

After proofreading, copy the result or download the UTF-8 text file. Open the saved file and compare the first and last lines, the example numbers and any uncommon Chinese characters. A successful download only confirms that a file was saved; this quick check confirms that it contains the text you intended to keep.

Keep an unchanged checked transcription if you also need a simplified or traditional version. Use the simplified and traditional Chinese converter on a copy, then review names and context-dependent wording. This separates what the source actually printed from a version adapted for another reader.

When pasting into a spreadsheet, set code columns to text before inserting values such as 0018. Keep quantities and their units together until you have deliberately split them into fields. Save the source image beside your working notes when appropriate, so that the next person can resolve a doubtful character without starting the task again.