1. Upload the image
Drop a PNG/JPG/WEBP into the Upload table image zone. The image is never uploaded — tesseract.js runs the Korean (kor) + English (eng) models in a WebAssembly engine inside your browser. On first use the engine and language data are fetched from the jsDelivr CDN (Korean 1.6MB + English 3.0MB + engine 4.1MB, about 8.6MB compressed); language data is cached in IndexedDB thereafter.
2. Capture for best accuracy
- Resolution — 1600px on the long edge is ideal. We upscale automatically, but crisp originals help.
- Perspective — keep the camera perpendicular to the table. This tool does not deskew, and the Tesseract docs note that skew significantly reduces line-segmentation quality.
- Contrast — we grayscale + histogram-stretch internally; still, avoid dark or patterned backgrounds.
- Crop margin — leave some padding around the table; the Tesseract docs warn that text cropped with no border can cause problems.
- Grid lines — optional. We do not use lines; row/column split is derived purely from word bounding boxes.
3. How rows and columns are inferred
We take each word's vertical centre to cluster into rows, and use 1-D clustering on left-edge X values to find column starts. Words in the same row falling in the same column stripe are joined with spaces. Heavy skew or multi-line cells degrade accuracy; use the preview to edit cells or delete mis-split rows.
4. Edit & export
- Edit a cell — double-click any cell to type; Enter or blur to save.
- Select rows — checkbox on the left; then "Delete selected rows" to bulk-remove.
- Add a row — appends an empty row with the same column count.
- Download XLSX — single-sheet "OCR" workbook.
- Download CSV — UTF-8 with BOM so Korean/Japanese headers render in Excel.
5. Korean-aware post-processing
Each cell runs through quick clean-ups:
12,345 원→12,345원(no-space before the common currency/unit suffixes 원 · 만원 · 억원 · 달러 · 엔 · 위안 · % · 개 · 건 · 명 · 회)- Fullwidth comma
,→, - Standalone noise cells like
|/·removed (often grid-line artefacts)
These clean-ups cannot recover text the underlying OCR already lost. If accuracy is poor, retake the image at a higher resolution — or edit cells manually.
6. Numbers & dates
- Thousand separators stay as-is. To treat as numbers in Excel, use Text-to-Columns → remove comma, or paste as values.
- Dates (e.g. 2025-03-01) come through as strings. Convert with
=DATEVALUE(A2)or cell format "Date". - When Korean unit suffixes are mixed in with numeric values, splitting number and unit into separate columns generally helps downstream analysis.
7. Performance
Processing takes from a few seconds to tens of seconds depending on your device and image size. The first run also downloads about 8.6MB of engine and language data. Very large images hit memory limits — crop to just the table region for faster and more accurate recognition.
8. Privacy
The site is static hosting (S3 + CloudFront) — no upload endpoint exists. Open DevTools → Network and run OCR: no request carries your image; you only see the first-run engine/language downloads from cdn.jsdelivr.net and analytics pings. Some organisations block third-party CDN scripts, so check your security policy first.