Recursive AI

Convert a scanned page to Markdown in C#

OneOcr is a free, open-source OCR library for .NET from Recursive AI. MIT or Apache-2.0, no native dependencies.

The zip includes page.png, a stock report.

Terminal
dotnet add package RecursiveAI.OneOcr

Requires .NET 10 (net10.0); .NET 9 and earlier aren’t supported.

C#
using OneImage;
using OneOcr;
var engine = OcrEngine.LoadBundled();
File.WriteAllText("page.md",
    Image.Load("page.png").ToMarkdown(engine));

Unzip the example, then dotnet run.

Writes page.md (no console output):

MARKDOWN
# STOCK REPORT
 
## Items ready for delivery.
 
| Product | Quantity |
| --- | --- |
| Paper | 12 |
| Pens | 24 |
 
Please check all quantities.
 
## Store paper in a dry room.
 
Contact the office for supplies.

The full program

The input has a large “STOCK REPORT” title, “Items ready for delivery.”, a two-column Product/Quantity table containing Paper/12 and Pens/24, and three closing instructions. The output above shows all of those inputs, including the extra headings inferred from the layout.

C#
using OneImage;
using OneOcr;
var engine = OcrEngine.LoadBundled();
var image = Image.Load("page.png");
string markdown = image.ToMarkdown(engine);
File.WriteAllText("page.md", markdown);
Console.Write(markdown);

The full program also prints the Markdown shown above. The OneImage namespace comes from a dependency installed with RecursiveAI.OneOcr.

Unreadable image bytes raise OcrFormatException; a readable image with no recognized text produces an empty Markdown string.

Variations

Keep running headers and footers, which are normally omitted.

C#
using OneImage;
using OneOcr;
var engine = OcrEngine.LoadBundled();
var image = Image.Load("page.png");
Console.Write(image.ToMarkdown(engine,
    new MarkdownOptions { IncludeHeadersAndFooters = true }));

On this sample, the output is identical to the Markdown printed above; no running header or footer was detected.

The short answer

Load the scan as a OneImage Image, create OcrEngine.LoadBundled(), and call image.ToMarkdown(engine). Save the returned string with File.WriteAllText. The full example writes a Markdown file, not an image embedded in a Markdown document. Its pipe table contains the recognized cells, so you can edit the quantities without modifying the original scan.

This is a local, rule-based Markdown writer over OCR results. It does not use a language model to rewrite the document. The Markdown specification describes which layout features are represented and which are not.

Can an image be converted to Markdown?

Yes, but the conversion has two distinct parts: recognize the words and their layout, then serialize that layout as Markdown. The sample demonstrates headings, paragraphs and a table. Its title becomes # STOCK REPORT; the first table row becomes the Markdown header row, followed by the separator and data rows. The writer uses the first row as a header convention, not as a guarantee that the document’s author intended it as one.

Review the inferred structure as well as the words. In this run, “Items ready for delivery.” and “Store paper in a dry room.” become second-level headings. They are ordinary sentences in the generated input. The program has not changed their wording, but its layout interpretation is imperfect. If your application needs authoritative section boundaries, review the result instead of treating every # marker as a verified document label.

A table cell spanning several rows or columns places its text at the top-left position; the covered positions stay empty. This keeps cell text available, but a Markdown pipe table cannot reproduce merged-cell geometry. Text containing a pipe is escaped so that it does not accidentally start another column.

The writer does not emit bold or italic emphasis, links, embedded images, code blocks, footnotes or math. A figure contributes text recognized inside it, not a saved image. Multi-column material follows the page’s reading order rather than retaining a side-by-side page layout. Choose this output for editable text with basic structure, not for a visual copy of the scan.

How to convert image to text using OCR?

The same recognition engine can produce plain text, structured pages or Markdown. Use plain text when all you need is the wording. Use Markdown when a table or heading should survive into an editor. Keep the structured page model when you need word coordinates, confidence, diagnostics or the original table spans. Markdown deliberately leaves those details out.

Start with an upright, legible page image and compare the output with it. A recognized amount in a table remains recognized text; Markdown does not validate the number or its meaning. The sample’s simple generated image lets you check the expected cells directly. Real scans can introduce blur, noise and skew that change the transcription or structure.

Running headers and footers are omitted by default. The variation keeps blocks assigned those roles, but it does not invent a header if none was detected. Page-number removal is a separate Markdown option. For multipage images, the image extension reads every frame; use OcrPages when you need to inspect each page and its diagnostics separately before exporting.

All oneocr recipes