FieldCure.DocumentParsers.Ocr 1.2.0

FieldCure.DocumentParsers.Ocr

Tesseract OCR fallback for scanned PDFs.

This package plugs into FieldCure.DocumentParsers by registering an OcrPdfParser for .pdf — when PdfPig yields no text layer for a page, the page is rendered at 300 DPI (via PDFium) and recognized with Tesseract.

⚠ Platform: Windows only (x64 + arm64). Native Tesseract binaries are bundled for both architectures: win-x64 ships Tesseract 5.0 / Leptonica 1.82.0 (redistributed from upstream Tesseract 5.2.0), and win-arm64 ships Tesseract 5.5.2 / Leptonica 1.87.0 built from source via vcpkg and Authenticode-signed by FieldCure. The correct-arch DLLs are selected at consumer build time. The assembly is marked [SupportedOSPlatform("windows")] — cross-platform consumers will see a CA1416 warning at compile time. Linux / macOS support is planned for a future release.

If you only need text from PDFs with an embedded text layer, use the core FieldCure.DocumentParsers package (pure managed, fully cross-platform).

Install

dotnet add package FieldCure.DocumentParsers.Ocr

Quick Start

using FieldCure.DocumentParsers;
using FieldCure.DocumentParsers.Ocr;

// Registers OcrPdfParser with a Tesseract engine. Dispose the engine at shutdown.
using var ocr = DocumentParserFactoryOcrExtensions.AddOcrSupport();

// Use the factory as usual — scanned pages are OCR'd automatically.
var parser = DocumentParserFactory.GetParser(".pdf")!;
var text = parser.ExtractText(File.ReadAllBytes("scanned.pdf"));

Custom engine

using var myEngine = new TesseractOcrEngine(maxPoolSize: 8);
DocumentParserFactoryOcrExtensions.AddOcrSupport(myEngine);

Implement IOcrEngine to use a different OCR backend.

How It Works

1. PdfPig extracts text for each page (same pipeline as the core package).
2. If a page yields less than 5% non-whitespace or fewer than 10 meaningful chars, the page is rendered at 300 DPI via PDFium.
3. The rendered image is fed to the IOcrEngine.
4. Korean output is post-processed to remove spurious inter-character spaces.

Included Languages

• English (eng.traineddata)
• Korean (kor.traineddata)

Languages are auto-discovered from embedded traineddata files.

Thread Safety

TesseractOcrEngine uses an engine pool (default size: min(ProcessorCount, 4)).

Architecture support

The Tesseract C ABI is additive across 5.0 → 5.5 (no symbol removals or signature changes), so the wrapper's [DllImport] surface is fully compatible. build/FieldCure.DocumentParsers.Ocr.targets selects the correct-arch DLL at consumer build time by detecting host RID via $(NETCoreSdkRuntimeIdentifier), with $(Platform), $(RuntimeIdentifier), $(PlatformTarget) as override paths for explicit-arch cross builds.

PackAsTool consumers (dnx, dotnet tool)

A single tool nupkg may be fetched and run by dnx on either x64 or ARM64, but it's packed once on a single CI host. From v1.2.0 on, PackAsTool consumers get both native trees inside the tool nupkg:

tools/<tfm>/any/x64/                 -- x64 Tesseract + Leptonica
tools/<tfm>/any/arm64-platform/x64/  -- ARM64 Tesseract + Leptonica

At runtime, NativeLibraryBootstrap (called from TesseractOcrEngine's constructor) inspects RuntimeInformation.ProcessArchitecture and, on ARM64, points the wrapper's LibraryLoader.CustomSearchPath at the arm64-platform/ subfolder. The wrapper's hard-coded <base>\x64\ lookup then resolves the ARM64 binaries. The bootstrap is a silent no-op for library consumers (where the chosen-arch DLL already sits in <base>\x64\ directly).

Related Packages

• FieldCure.DocumentParsers — Core text extraction
• FieldCure.DocumentParsers.Imaging — Page image rendering

License

MIT — Copyright (c) 2026 FieldCure Co., Ltd.