OneOcr reads sideways scans and straightens tilted text lines before recognition.
dotnet add package RecursiveAI.OneOcrRequires .NET 10 (net10.0); .NET 9 and earlier aren’t supported.
using OneOcr; var engine = OcrEngine.LoadBundled(); var page = engine .OcrPages("turned.png").Single(); Console.WriteLine(page.ToText());
Unzip the example, then dotnet run.
turned.png is a sideways train note in the download. The snippet prints its text upright:
Morning train to the coast Meet at the station Bring a map and water Return before sunset
The full program
The download has three generated PNGs with those same four lines: turned.png (90° clockwise), tilted.png (8° clockwise), and steep32.png (32° clockwise). The program checks the text, the input dimensions, and the first word’s position. On the sideways sample, “Morning” is near the right edge of the input, not the top left of a new upright image.
using System.Globalization; using OneOcr; CultureInfo.CurrentCulture = CultureInfo.InvariantCulture; var engine = OcrEngine.LoadBundled(); var samples = new[] { (File: "turned.png", Width: 275, Height: 560, Turn: 90, Skew: 0, Searched: false), (File: "tilted.png", Width: 594, Height: 351, Turn: 0, Skew: 8, Searched: false), (File: "steep32.png", Width: 622, Height: 531, Turn: 0, Skew: 20, Searched: true) }; const string expected = "Morning train to the coast\nMeet at the station\nBring a map and water\nReturn before sunset"; foreach (var sample in samples) { var page = engine.OcrPages(sample.File).Single(); var text = page.ToText().ReplaceLineEndings("\n").Trim(); var word = page.Words.First(); bool searched = page.Diagnostics.Any(d => d.Code == OcrDiagnosticCodes.AngleSearched); if (text != expected || page.Orientation != sample.Turn || Math.Abs(page.SkewDegrees - sample.Skew) > 0.2 || searched != sample.Searched || page.Width != sample.Width || page.Height != sample.Height || word.Quad is null || word.Box.Left < 0 || word.Box.Top < 0 || word.Box.Right > sample.Width || word.Box.Bottom > sample.Height || (sample.Turn == 90 && word.Box.Left < sample.Width / 2)) throw new InvalidOperationException($"Unexpected OCR result for {sample.File}."); Console.WriteLine($"{sample.File}: turn {page.Orientation}°, skew {page.SkewDegrees:F1}°"); Console.WriteLine(text); Console.WriteLine($"{word.Text} at ({word.Box.Left:F0}, {word.Box.Top:F0}) on {page.Width}×{page.Height}; quad: yes"); Console.WriteLine($"Angle-search diagnostic: {searched}"); }
dotnet run prints:
turned.png: turn 90°, skew 0.0° Morning train to the coast Meet at the station Bring a map and water Return before sunset Morning at (211, 40) on 275×560; quad: yes Angle-search diagnostic: False tilted.png: turn 0°, skew 8.0° Morning train to the coast Meet at the station Bring a map and water Return before sunset Morning at (70, 37) on 594×351; quad: yes Angle-search diagnostic: False steep32.png: turn 0°, skew 20.0° Morning train to the coast Meet at the station Bring a map and water Return before sunset Morning at (146, 48) on 622×531; quad: yes Angle-search diagnostic: True
For the 32° sample, skew 20.0° is the winning reading’s reported skew, not the image’s actual tilt.
If the page has no clear text direction, orientation detection reads it as given. Check the returned text before using it as a transcript.
Variations
Turn off the extra angle search to compare with the first reading. Replace the OcrPages call at the top with these lines; DetectOrientation and Deskew stay on.
var options = new OcrOptions { SearchAngle = false }; var page = engine.OcrPages("steep32.png", options).Single(); Console.WriteLine(page.SkewDegrees.ToString("F1", System.Globalization.CultureInfo.InvariantCulture));
On steep32.png this reports 12.0 instead of 20.0.
Related recipes
The short answer
Call OcrPages(path) to read a rotated or skewed image in C#. DetectOrientation, Deskew, and SearchAngle are on by default. Read the result with page.ToText(). If you’re highlighting matches on the scan, use the word boxes or corner coordinates from the same result. You don’t need to rotate those coordinates back yourself.
How do I correct the orientation of a scanned document for OCR?
Orientation detection chooses among upright, sideways, and upside down: 0°, 90°, 180°, and 270°. OcrPage.Orientation tells you how far the content was turned clockwise from upright. On turned.png it’s 90°, meaning the tops of the lines point right. It isn’t an instruction to turn the image another 90°.
The correction is for reading. The OCR result doesn’t save a new image. In the sideways sample, Width and Height still describe the 275 × 560 input. Box is a rectangle with horizontal and vertical sides around the word. Quad gives the four corners in the text’s own order: top left, top right, bottom right, bottom left. Use the rectangle for a rough highlight, or the four corners if the highlight should follow the tilted text. See the page geometry docs.
How do I deskew a tilted image for OCR?
Deskew corrects a smaller tilt, such as a page that went into a scanner at an angle. It straightens text lines tilted by up to about 15°; tilts under about 0.3° are left alone. SkewDegrees reports the tilt removed before reading, positive for lines descending to the right. That’s why the 8° image can have Orientation of 0° and still be straightened.
When the first reading is weak, SearchAngle can try other whole-page angles and keep a reading with more confidently read text. The diagnostic OcrDiagnosticCodes.AngleSearched marks when a different reading won; its absence doesn’t establish that no angles were tried. In the steep sample, the winning reported skew was 20° even though the image was made with a 32° turn. Treat this as a reading choice, not a measurement of the camera angle or a promise that every image at 32° works. See how OneOcr finds and maps angles.