anyocr

One unified API over every major OCR engine — from Tesseract to vision LLMs.

anyocr gives you a single anyocr.read() call that extracts text from images and documents using whichever OCR engine you have installed. It auto-selects the best available backend by priority (Surya > EasyOCR > PaddleOCR > Tesseract > Vision-LLM), applies a smart preprocessing pipeline (auto-rotate, deskew, contrast enhancement, binarization), and returns structured results with bounding boxes, confidence scores, and reading order.

Built by Viet-Anh Nguyen at NRL.ai.

Why anyocr?

• One-liner API — anyocr.read("scan.png") just works with any installed backend
• Plugin architecture — Register new OCR engines via @register_backend
• Local-first — Surya, EasyOCR, Paddle, Tesseract all run on your machine
• Minimal core deps — Only pillow and numpy; every OCR engine is an optional extra
• Production-ready — Auto-preprocessing, structured dataclass results, batch inference

Installation

pip install anyocr

For optional backends:

pip install anyocr[surya]        # Surya OCR — SOTA open source, 90+ languages
pip install anyocr[easyocr]      # EasyOCR — CRNN models, 80+ languages
pip install anyocr[paddle]       # PaddleOCR — strong Asian languages
pip install anyocr[tesseract]    # Tesseract via pytesseract (needs tesseract binary)
pip install anyocr[vlm]          # Vision-LLM via anyllm (GPT-4V, Claude, Gemini)
pip install anyocr[all]          # everything

Python 3.8+ supported (tested on 3.8, 3.9, 3.10, 3.11, 3.12, 3.13)

Quick Start

import anyocr

# 1. Auto-selects the best installed backend (Surya > EasyOCR > Paddle > Tesseract > VLM)
result = anyocr.read("receipt.png")
print(result.text)                       # full extracted text
for line in result.lines:
    print(line.text, line.confidence, line.bbox)

# 2. Force a specific backend
text = anyocr.read("chinese.jpg", backend="paddle", lang="ch").text

# 3. Use a vision LLM for hard cases (tables, handwriting)
text = anyocr.read("handwritten.jpg", backend="vlm", model="gpt-4o").text

Models & Methods

Supported backends (auto-selected by priority)

You can change the priority or force a backend via anyocr.read(..., backend="easyocr") or anyocr.set_priority(["paddle", "surya"]).

Preprocessing pipeline

Applied automatically (can be disabled per call):

1. EXIF orientation fix — rotate based on metadata
2. Auto-rotate — detect 90/180/270 rotation via text-line angle histogram
3. Deskew — Hough-transform-based angle correction (<= 15 degrees)
4. Contrast enhancement — CLAHE (adaptive histogram equalization)
5. Binarization — adaptive threshold for low-quality scans (opt-in)
6. Denoise — bilateral filter for scanned documents (opt-in)

Result dataclasses

@dataclass
class OCRLine:
    text: str
    confidence: float
    bbox: tuple[float, float, float, float]   # x1, y1, x2, y2
    polygon: list[tuple[float, float]] | None  # 4-point quad if supported

@dataclass
class OCRResult:
    text: str                  # joined full text in reading order
    lines: list[OCRLine]
    backend: str               # which backend produced this result
    language: str | None

API Reference

CLI Usage

anyocr read receipt.png
anyocr read scan.jpg --backend paddle --lang ch
anyocr read-pdf document.pdf --out text.txt
anyocr list-backends

Examples

OCR an entire PDF and save as text

import anyocr

# Rasterizes each page and runs the auto-selected backend
result = anyocr.read_pdf("report.pdf")
with open("report.txt", "w") as f:
    for page_num, page in enumerate(result.pages, 1):
        f.write(f"=== Page {page_num} ===\n{page.text}\n\n")

Combine preprocessing with a specific backend

import anyocr

# Run the preprocessing pipeline explicitly before OCR
cleaned = anyocr.preprocess("noisy_scan.jpg", ops=["deskew", "clahe", "binarize"])
result  = anyocr.read(cleaned, backend="tesseract", lang="eng")
print(result.text)

Compare two backends on the same image

import anyocr

for backend in ["surya", "easyocr", "paddle"]:
    r = anyocr.read("test.jpg", backend=backend)
    print(f"{backend}: {r.text[:80]}... (avg conf {r.mean_confidence():.2f})")

License

MIT (c) Viet-Anh Nguyen