natocr 1.6.0

Native OCR library using platform-specific frameworks (macOS Vision, Windows Runtime OCR)

natocr

natocr (native ocr) is a small Python wrapper around the OCR engines that already ship with macOS and Windows: Vision framework on macOS and Windows Runtime OCR on Windows.

These built-in engines are generally faster, more efficient, and more accurate than third-party alternatives like Tesseract. natocr makes reaching for them painless via one clean Python API instead of wrangling with Objective-C bridges or WinRT async plumbing.

Install

pip install natocr

# for JPEG XL, JPEG XR & DjVu support
pip install natocr[extras]

The right native backend (Vision on macOS, Windows Runtime OCR on Windows) is pulled in automatically for your platform - no OS-specific install command to pick.

Quick start

from natocr import OCR

ocr = OCR()                    # defaults to english
pages = ocr.recognize("invoice.png")   # one OCRResult per page

print(pages[0].text)

Invoice #1042 Total $58.20 Thank you!

recognize() always returns a list of OCRResult - one per page. Most images are a single page, so you'll often just read pages[0]; multi-page documents (DjVu, TIFF, GIF) give one result per page (see Multi-page documents).

Confidence Scores and Bounding Boxes

Beyond the flat .text, each OCRResult carries a per-detection breakdown with bounding boxes and (on macOS) confidence scores:

page = ocr.recognize("receipt.png")[0]   # first (often only) page

print(page.confidence)            # average confidence, or None if unavailable

for element in page.elements:
    box = element.bounds.bounds   # (x, y, width, height) in pixels
    print(f"{element.text!r} @ {box} conf={element.confidence}")

0.93
'Acme Coffee' @ (24.0, 18.0, 180.0, 32.0) conf=0.97
'Latte' @ (24.0, 70.0, 96.0, 28.0) conf=0.95
'$4.50' @ (220.0, 70.0, 80.0, 28.0) conf=0.88

Lines and Words

There's also convenience views for grouping a page by reading order:

page.lines      # ['Acme Coffee', 'Latte $4.50']  - elements grouped into lines
page.words      # list of TextElement with non-empty text

Detection Language

Pick a different recognition language, and inspect what the current platform supports:

ocr = OCR(language="fr")
print(ocr.platform)               # 'darwin' or 'win32'
print(ocr.supported_languages)    # ['en-US', 'fr-FR', 'de-DE', ...]

The supported set is decided by the OS and queried live, so supported_languages always reflects the current machine. On macOS it's Vision's built-in set for your macOS version; on Windows it's whatever OCR language packs are installed. See the Usage guide for the full list and how to add Windows language packs.

Alternative Inputs

recognize() accepts more than file paths - hand it whatever you already have in memory:

from PIL import Image
import numpy as np

ocr.recognize("page.png")              # a file path
ocr.recognize(Image.open("page.png"))  # a PIL image
ocr.recognize(np.array(image))         # a numpy array (e.g. from OpenCV)
ocr.recognize(open("page.png", "rb").read())  # raw image bytes

Supported File Types

Images are decoded with Pillow, so any raster format Pillow can open works as an input file or byte string. HEIC/HEIF decoding (and AVIF) is provided by the bundled pillow-heif, so iPhone photos work with no extra setup. JPEG XL, JPEG XR, and DjVu need extra decoders - install them with pip install natocr[extras] (see Optional formats below).

Format	Extensions	Notes
AVIF	.avif	AV1-based, decoded via the bundled pillow-heif
BMP	.bmp	uncompressed bitmap
DjVu	.djvu, .djv	scanned documents; multi-page (needs natocr[extras] + the djvulibre system library)
GIF	.gif	multi-page - one result per frame
HEIC/HEIF	.heic, .heif, .hif	iPhone photos and screenshots
JPEG	.jpg, .jpeg	great for photos of documents
JPEG 2000	.jp2, .j2k, .jpf, .jpx	wavelet-based, decoded natively by Pillow
JPEG XL	.jxl	modern successor to JPEG (needs natocr[extras])
JPEG XR / HD Photo	.jxr, .wdp, .hdp	Microsoft HD Photo (needs natocr[extras])
PCX	.pcx	legacy PC Paintbrush, common in old scan archives
PNG	.png	recommended - lossless
PPM/PGM	.ppm, .pgm	netpbm bitmaps
TIFF	.tif, .tiff	common for scans; multi-page
WebP	.webp	modern lossy/lossless

Optional formats (JPEG XL, JPEG XR, DjVu)

These are optional because their decoders are extra dependencies. Install them with:

pip install natocr[extras]

That pulls in pillow-jxl-plugin for .jxl, imagecodecs for .jxr/.wdp/.hdp, and python-djvulibre for .djvu/.djv. Once installed they decode through the same recognize() call as every other format - no extra code. Without the extra, the rest of the formats above (including JPEG 2000) keep working unchanged.

DjVu also needs the system djvulibre library that python-djvulibre builds against:

brew install djvulibre         # macOS
sudo apt install libdjvulibre-dev   # Debian/Ubuntu

On Windows, install DjVuLibre so its DLLs land on PATH (the wheel links against it).

[!NOTE] Support degrades gracefully: if natocr[extras] or the djvulibre library isn't present, DjVu just isn't registered and opening a .djvu raises Pillow's usual UnidentifiedImageError. Every other format keeps working - nothing else breaks.

Multi-page documents

recognize() reads every page and returns one OCRResult per page, in order. The formats that can carry more than one page are DjVu, multi-page TIFF, and animated GIF:

for i, page in enumerate(ocr.recognize("scan.djvu"), start=1):
    print(f"--- page {i} ---")
    print(page.text)

Single-page inputs (PNG, JPEG, ...) come back as a one-element list, so the same loop works for everything - or just grab recognize(...)[0].

In addition to file paths, recognize() accepts these in-memory types:

Input type	Example
str (file path)	ocr.recognize("page.png")
PIL.Image.Image	ocr.recognize(Image.open("page.png"))
numpy.ndarray	ocr.recognize(np.array(image))
bytes (encoded image)	ocr.recognize(data)

[!NOTE] Only DjVu, TIFF, and GIF carry multiple pages here. PDFs aren't decoded directly - rasterize a page to one of the formats above first (e.g. with pdf2image or pymupdf).

Testing

Install the dev dependencies (in a virtualenv), then run the suite. The tests mock the native macOS Vision and Windows Runtime backends, so they run anywhere without those frameworks installed.

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Run everything with coverage (coverage is wired up in pyproject.toml, so plain pytest already reports it):

pytest

Other handy invocations:

# run a single test file
pytest tests/test_models.py

# run one test by name
pytest -k test_lines_groups_close_y_into_single_line

# verbose output
pytest -v

Coverage reports land in the terminal, in htmlcov/index.html, and in coverage.xml.