Tesseract OCR for OpenHarmony / HarmonyOS

This directory ports the upstream Tesseract OCR engine (libtesseract) and its dependency chain to OpenHarmony (OHOS), and ships a DevEco Studio demo app that performs on-device OCR through a NAPI bridge.

The Tesseract C++ sources are unmodified — the whole port is done through the OpenHarmony NDK toolchain plus the build glue in this folder.

ohos/
├── scripts/
│   ├── env.sh              # locates the OHOS NDK, sets arch/triple/clang/sysroot
│   ├── fetch-sources.sh    # downloads zlib / libjpeg-turbo / libpng / leptonica
│   ├── build-deps.sh       # cross-compiles the dependency chain (static, PIC)
│   ├── build-tesseract.sh  # cross-compiles libtesseract.a + installs headers
│   └── build-all.sh        # one-shot: fetch + deps + tesseract
├── prebuilt/<arch>/        # OUTPUT: { include/, lib/*.a } — the OCR SDK (gitignored)
└── demo/TessOcrDemo/       # DevEco Studio app that calls the engine via NAPI

1. Prerequisites

• macOS / Linux with the OpenHarmony Native NDK installed (this port was built against API 21 / native 6.0.1.112, clang-15). The scripts auto-detect it from HOS_SDK_HOME, OHOS_NATIVE_HOME, or the DevEco Studio bundle.
• cmake (≥ 3.16) and ninja on PATH.

2. Build the engine

cd ohos
# default target is arm64-v8a (real devices)
bash scripts/build-all.sh
# or for the emulator:
OHOS_ARCH=x86_64 bash scripts/build-all.sh

Result in ohos/prebuilt/arm64-v8a/:

Library	Purpose
libtesseract.a	OCR engine
libleptonica.a	image I/O used by Tesseract
libpng16.a, libjpeg.a, libz.a	image codecs

All objects are aarch64-linux-ohos ELF, position-independent, ready to be linked into an OHOS native module (.so).

3. Dependency chain

zlib ──► libpng ─┐
libjpeg-turbo ───┼─► leptonica ──► libtesseract
                 ┘

TIFF/WEBP/GIF/OpenJPEG, the training tools (ICU/Pango/Cairo), libcurl and libarchive are disabled — none are needed for on-device recognition, which keeps the cross-build self-contained. PNG + JPEG cover the common image inputs; re-enable TIFF by building libtiff first and flipping ENABLE_TIFF/DISABLE_TIFF.

4. What the port had to handle

The engine is portable C++17; OHOS uses a clang + musl + CMake toolchain very close to the Android NDK that Tesseract already supports. The only friction was in the build, not the code:

• CMAKE_SYSTEM_NAME=OHOS — the toolchain sets UNIX=TRUE, so Tesseract's generic Unix path is used (and the Android-only CpuFeaturesNdkCompat dependency is skipped).
• SIMD — on __aarch64__, NEON is always-on at compile time, so no runtime getauxval/android_getCpuFamily probing is required.
• libpng cross-build — its pnglibconf generator invokes clang directly and drops the toolchain target, so the per-arch musl include dir is re-injected (see build-deps.sh).
• Leptonica TIFF probe — Tesseract runs a try_run() to detect TIFF in Leptonica, which is illegal when cross-compiling; pre-seeding LEPT_TIFF_RESULT=1 skips it (see build-tesseract.sh).

5. Demo app

See demo/TessOcrDemo/README.md. It loads a bundled sample image + eng.traineddata, runs OCR off the UI thread via a Promise-returning NAPI call, and shows the recognized text.