决策树

Is accuracy critical and budget available?
├─ YES → Google Vision API or AWS Textract
└─ NO → Local solution
    ├─ CJK (Chinese/Japanese/Korean) or tables? → PaddleOCR
    ├─ General photos or multiple languages? → EasyOCR
    ├─ Simple printed English docs? → Tesseract
    └─ PDF documents with structure? → Docling or Surya

Python 实现

Tesseract (pytesseract)

import pytesseract
from PIL import Image
import cv2
import numpy as np

def extract_text_tesseract(image_path: str, lang: str = "eng") -> str:
    """Extract text using Tesseract. Best for clean printed documents."""
    image = Image.open(image_path)

    # Config: --psm 6 = assume uniform block of text
    config = "--psm 6 --oem 3"
    text = pytesseract.image_to_string(image, lang=lang, config=config)
    return text.strip()

def extract_with_confidence(image_path: str) -> list[dict]:
    """Extract text with bounding boxes and confidence scores."""
    image = Image.open(image_path)
    data = pytesseract.image_to_data(image, output_type=pytesseract.Output.DICT)

    results = []
    for i, word in enumerate(data["text"]):
        if word.strip() and int(data["conf"][i]) > 30:
            results.append({
                "text": word,
                "confidence": data["conf"][i],
                "bbox": {
                    "x": data["left"][i],
                    "y": data["top"][i],
                    "width": data["width"][i],
                    "height": data["height"][i],
                }
            })
    return results

# Install: pip install pytesseract pillow
# System: apt install tesseract-ocr (Linux) / brew install tesseract (Mac)

EasyOCR

import easyocr
from pathlib import Path

def extract_text_easyocr(
    image_path: str,
    languages: list[str] = ["en"],
    detail: bool = False
) -> str | list:
    """
    Extract text using EasyOCR. Best for photos and multiple languages.
    languages: ['en'], ['en', 'es'], ['ch_sim', 'en'], etc.
    """
    reader = easyocr.Reader(languages, gpu=False)  # gpu=True if CUDA available
    results = reader.readtext(image_path)

    if not detail:
        # Return plain text sorted by vertical position
        results_sorted = sorted(results, key=lambda x: x[0][0][1])
        return "\n".join([text for _, text, conf in results_sorted if conf > 0.3])

    return [
        {
            "text": text,
            "confidence": round(conf, 3),
            "bbox": bbox,
        }
        for bbox, text, conf in results
    ]

# Install: pip install easyocr

PaddleOCR (best for CJK and tables)

from paddleocr import PaddleOCR
import json

def extract_text_paddle(
    image_path: str,
    lang: str = "en",  # "en", "ch", "japan", "korean", "es", etc.
    use_angle_cls: bool = True,
) -> str:
    """Extract text using PaddleOCR. Best for CJK and structured documents."""
    ocr = PaddleOCR(use_angle_cls=use_angle_cls, lang=lang, show_log=False)
    result = ocr.ocr(image_path, cls=True)

    lines = []
    if result and result[0]:
        # Sort by y position (top to bottom)
        items = sorted(result[0], key=lambda x: x[0][0][1])
        lines = [item[1][0] for item in items if item[1][1] > 0.3]

    return "\n".join(lines)

# Install: pip install paddlepaddle paddleocr

Google Vision API

from google.cloud import vision
import io

def extract_text_google_vision(image_path: str) -> dict:
    """
    Extract text using Google Vision API.
    Requires: GOOGLE_APPLICATION_CREDENTIALS env var set.
    """
    client = vision.ImageAnnotatorClient()

    with io.open(image_path, "rb") as image_file:
        content = image_file.read()

    image = vision.Image(content=content)

    # Full text detection (better for documents)
    response = client.document_text_detection(image=image)
    document = response.full_text_annotation

    return {
        "text": document.text,
        "pages": [
            {
                "blocks": [
                    {
                        "text": " ".join(
                            symbol.text
                            for para in block.paragraphs
                            for word in para.words
                            for symbol in word.symbols
                        ),
                        "confidence": block.confidence,
                    }
                    for block in page.blocks
                ]
            }
            for page in document.pages
        ]
    }

# Install: pip install google-cloud-vision

AWS Textract (best for forms and invoices)

import boto3
import json

def extract_text_textract(image_path: str, region: str = "us-east-1") -> dict:
    """
    Extract text, forms, and tables using AWS Textract.
    Handles key-value pairs and structured tables automatically.
    """
    client = boto3.client("textract", region_name=region)

    with open(image_path, "rb") as f:
        image_bytes = f.read()

    response = client.analyze_document(
        Document={"Bytes": image_bytes},
        FeatureTypes=["TABLES", "FORMS"]
    )

    # Extract raw text
    blocks = response["Blocks"]
    lines = [b["Text"] for b in blocks if b["BlockType"] == "LINE"]

    # Extract key-value pairs (forms)
    key_values = {}
    key_map = {b["Id"]: b for b in blocks if b["BlockType"] == "KEY_VALUE_SET" and "KEY" in b.get("EntityTypes", [])}
    value_map = {b["Id"]: b for b in blocks if b["BlockType"] == "KEY_VALUE_SET" and "VALUE" in b.get("EntityTypes", [])}

    for key_block in key_map.values():
        key_text = _get_text_from_block(key_block, blocks)
        for rel in key_block.get("Relationships", []):
            if rel["Type"] == "VALUE":
                for val_id in rel["Ids"]:
                    if val_id in value_map:
                        val_text = _get_text_from_block(value_map[val_id], blocks)
                        key_values[key_text] = val_text

    return {
        "text": "\n".join(lines),
        "form_fields": key_values,
    }

def _get_text_from_block(block, all_blocks):
    word_ids = []
    for rel in block.get("Relationships", []):
        if rel["Type"] == "CHILD":
            word_ids.extend(rel["Ids"])

    block_map = {b["Id"]: b for b in all_blocks}
    words = [block_map[wid]["Text"] for wid in word_ids if wid in block_map and block_map[wid]["BlockType"] == "WORD"]
    return " ".join(words)

# Install: pip install boto3

图像预处理

预处理是影响 OCR 准确率的首要因素。运行 OCR 前务必应用。

import cv2
import numpy as np
from PIL import Image, ImageEnhance, ImageFilter

def preprocess_for_ocr(image_path: str, output_path: str = None) -> np.ndarray:
    """
    Full preprocessing pipeline for maximum OCR accuracy.
    Apply selectively based on image type.
    """
    img = cv2.imread(image_path)

    # 1. Convert to grayscale
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    # 2. Resize if too small (OCR works better at 300+ DPI)
    height, width = gray.shape
    if width < 1000:
        scale = 2000 / width
        gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)

    # 3. Deskew (fix rotation)
    gray = deskew(gray)

    # 4. Denoise
    denoised = cv2.fastNlMeansDenoising(gray, h=10)

    # 5. Binarization (choose one based on lighting)
    # Option A: Otsu (uniform lighting)
    _, binary = cv2.threshold(denoised, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

    # Option B: Adaptive (uneven lighting, shadows)
    # binary = cv2.adaptiveThreshold(denoised, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
    #                                 cv2.THRESH_BINARY, 11, 2)

    # 6. Morphological cleanup (remove noise dots)
    kernel = np.ones((1, 1), np.uint8)
    cleaned = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)

    if output_path:
        cv2.imwrite(output_path, cleaned)

    return cleaned

def deskew(image: np.ndarray) -> np.ndarray:
    """Correct image rotation using projection analysis."""
    coords = np.column_stack(np.where(image > 0))
    angle = cv2.minAreaRect(coords)[-1]

    if angle < -45:
        angle = -(90 + angle)
    else:
        angle = -angle

    if abs(angle) < 0.5:  # Skip if nearly straight
        return image

    h, w = image.shape
    center = (w // 2, h // 2)
    M = cv2.getRotationMatrix2D(center, angle, 1.0)
    return cv2.warpAffine(image, M, (w, h), flags=cv2.INTER_CUBIC,
                          borderMode=cv2.BORDER_REPLICATE)

def enhance_contrast(image_path: str) -> Image.Image:
    """Enhance contrast using PIL - useful for faded text."""
    img = Image.open(image_path).convert("L")
    enhancer = ImageEnhance.Contrast(img)
    return enhancer.enhance(2.0)

# Install: pip install opencv-python pillow

预处理决策指南

PDF 转文本提取

import fitz  # PyMuPDF - for native text extraction
from pdf2image import convert_from_path  # for scanned PDFs
import pytesseract

def extract_pdf_text(pdf_path: str, ocr_fallback: bool = True) -> str:
    """
    Smart PDF extraction:
    - Uses native text layer if available (fast, accurate)
    - Falls back to OCR for scanned pages
    """
    doc = fitz.open(pdf_path)
    full_text = []

    for page_num, page in enumerate(doc):
        # Try native text extraction first
        text = page.get_text().strip()

        if text and len(text) > 50:
            full_text.append(text)
        elif ocr_fallback:
            # Scanned page — render and OCR
            pix = page.get_pixmap(dpi=300)
            img_path = f"/tmp/page_{page_num}.png"
            pix.save(img_path)

            ocr_text = pytesseract.image_to_string(img_path)
            full_text.append(ocr_text)

    doc.close()
    return "\n\n".join(full_text)

# Install: pip install PyMuPDF pdf2image pytesseract
# System: apt install poppler-utils (for pdf2image on Linux)

提取文本的后处理

import re
from difflib import SequenceMatcher

def clean_ocr_text(text: str) -> str:
    """Standard cleanup for OCR output."""
    # Remove non-printable characters
    text = re.sub(r"[^\x20-\x7E\n\t]", "", text)

    # Normalize whitespace
    text = re.sub(r" +", " ", text)
    text = re.sub(r"\n{3,}", "\n\n", text)

    # Fix common OCR misreads
    corrections = {
        r"\b0(?=[a-zA-Z])": "O",    # 0 misread as O before letter
        r"(?<=[a-zA-Z])0\b": "O",    # O misread as 0 after letter
        r"\bl\b": "I",               # lowercase l misread as I (context-dependent)
        r"rn": "m",                  # rn → m (common serif font error)
    }
    for pattern, replacement in corrections.items():
        text = re.sub(pattern, replacement, text)

    return text.strip()

def extract_structured_data(text: str) -> dict:
    """Extract common structured fields from OCR text."""
    patterns = {
        "email": r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b",
        "phone": r"[\+]?[(]?[0-9]{3}[)]?[-\s\.]?[0-9]{3}[-\s\.]?[0-9]{4,6}",
        "date": r"\b\d{1,2}[/-]\d{1,2}[/-]\d{2,4}\b",
        "amount": r"\$\s?\d+(?:,\d{3})*(?:\.\d{2})?",
        "url": r"https?://[^\s]+",
    }

    return {
        field: re.findall(pattern, text)
        for field, pattern in patterns.items()
    }

def merge_multiline_words(text: str) -> str:
    """Fix hyphenated words split across lines (common in PDFs)."""
    return re.sub(r"(\w)-\n(\w)", r"\1\2", text)

Node.js / TypeScript

// Using Tesseract.js (pure JS, no native deps needed)
import Tesseract from "tesseract.js";

async function extractText(imagePath: string, lang = "eng"): Promise<string> {
  const { data } = await Tesseract.recognize(imagePath, lang, {
    logger: () => {}, // suppress progress logs
  });
  return data.text.trim();
}

// With confidence filtering
async function extractWithConfidence(imagePath: string) {
  const { data } = await Tesseract.recognize(imagePath, "eng");

  return data.words
    .filter((word) => word.confidence > 70)
    .map((word) => ({
      text: word.text,
      confidence: word.confidence,
      bbox: word.bbox,
    }));
}

// Install: npm install tesseract.js



// Using Google Vision API from Node.js
import vision from "@google-cloud/vision";

const client = new vision.ImageAnnotatorClient();

async function extractTextCloud(imagePath: string): Promise<string> {
  const [result] = await client.documentTextDetection(imagePath);
  return result.fullTextAnnotation?.text ?? "";
}

// Install: npm install @google-cloud/vision

用于 OCR 的 Claude Vision API

当您需要结构化提取 + 理解时，使用 Claude 的视觉能力：

import anthropic
import base64
from pathlib import Path

def extract_with_claude(image_path: str, instruction: str = None) -> str:
    """
    Use Claude to extract and structure text from an image.
    Best when you need semantic understanding, not just raw text.
    """
    client = anthropic.Anthropic()

    image_data = base64.standard_b64encode(Path(image_path).read_bytes()).decode()
    ext = Path(image_path).suffix.lower()
    media_types = {".jpg": "image/jpeg", ".jpeg": "image/jpeg", ".png": "image/png", ".webp": "image/webp"}
    media_type = media_types.get(ext, "image/jpeg")

    prompt = instruction or (
        "Extract ALL text from this image exactly as it appears. "
        "Preserve the original structure, line breaks, and formatting. "
        "Return only the extracted text, nothing else."
    )

    message = client.messages.create(
        model="claude-opus-4-6",
        max_tokens=4096,
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "image",
                        "source": {
                            "type": "base64",
                            "media_type": media_type,
                            "data": image_data,
                        },
                    },
                    {"type": "text", "text": prompt},
                ],
            }
        ],
    )

    return message.content[0].text

# Example: structured invoice extraction
def extract_invoice(image_path: str) -> dict:
    result = extract_with_claude(
        image_path,
        instruction="""Extract all data from this invoice and return as JSON:
{
  "invoice_number": "",
  "date": "",
  "vendor": {"name": "", "address": "", "email": ""},
  "items": [{"description": "", "quantity": 0, "unit_price": 0, "total": 0}],
  "subtotal": 0,
  "tax": 0,
  "total": 0
}
Return only valid JSON, no explanation."""
    )
    import json
    return json.loads(result)

何时使用 Claude 与传统 OCR

按图像类型划分的准确率基准

*需要额外安装语言包

常见错误及修复

Tesseract 返回乱码

• 原因 : 图像太小或噪点太多
• 修复 : 放大 2 倍，应用去噪和二值化

EasyOCR 遗漏多列文本

• 原因 : 默认布局分析在多列上失败
• 修复 : 分别裁剪每一列并单独进行 OCR

PaddleOCR 在 CPU 上运行缓慢

• 原因 : 加载了大模型
• 修复 : 如果可用则使用 use_gpu=True，或对水平文本使用 use_angle_cls=False

边界框与文本不对齐

• 原因 : 图像在 OCR 前被旋转
• 修复 : 在预处理中应用 deskew()

云 API 在某些区域返回空结果

• 原因 : 对比度低或文字非常小
• 修复 : 预处理图像，提高 DPI，裁剪感兴趣区域

PDF 文本层编码错误

• 原因 : 非标准字体嵌入
• 修复 : 使用 fitz.Page.get_text("rawdict") 检查编码，或跳过并回退到 OCR

快速入门模板

最小化本地 OCR (Python)

pip install easyocr
python -c "import easyocr; r=easyocr.Reader(['en']); print('\n'.join([t for _,t,c in r.readtext('image.png') if c>0.3]))"

最小化云端 OCR (Node.js)

npm install tesseract.js
node -e "const T=require('tesseract.js'); T.recognize('image.png','eng').then(r=>console.log(r.data.text))"

批处理管道

from pathlib import Path
import easyocr

reader = easyocr.Reader(["en"], gpu=False)

def batch_ocr(folder: str, output_folder: str) -> None:
    Path(output_folder).mkdir(exist_ok=True)
    images = list(Path(folder).glob("*.{png,jpg,jpeg,tiff,bmp}"))

    for img_path in images:
        results = reader.readtext(str(img_path))
        text = "\n".join(t for _, t, c in results if c > 0.3)

        out_path = Path(output_folder) / f"{img_path.stem}.txt"
        out_path.write_text(text, encoding="utf-8")
        print(f"✓ {img_path.name} → {out_path.name}")

    print(f"\nProcessed {len(images)} images.")

batch_ocr("./images", "./output")

规则

• 在编写代码前，根据文档类型和准确率要求选择 OCR 引擎：Tesseract 用于本地/离线简单文档，EasyOCR 用于多语言手写体，云 API（Google Vision、AWS Textract）用于生产环境中对结构化文档的高准确率需求
• 对于非理想输入，在 Tesseract 和 EasyOCR 之前必须进行图像预处理（灰度转换、二值化、去倾斜）—— 跳过此步骤会导致准确率显著下降
• OCR 输出必须始终被视为未经验证的文本 —— 在将提取的值用于业务逻辑之前，应用后处理（正则表达式、字符串规范化）
• 在未与项目负责人确认数据隐私和合规性要求之前，切勿将敏感文档图像传递给云 OCR API
• 必须检查 OCR 引擎返回的置信度分数；低于项目定义阈值的结果必须标记为需要人工审核，而不是自动接受

每周安装次数

689

仓库

fearovex/claude-config

首次出现

2026年3月4日

安全审计

Gen Agent Trust HubPassSocketPassSnykFail

安装于

gemini-cli688

codex688

opencode688

amp687

github-copilot687

cursor687

Decision Tree

Is accuracy critical and budget available?
├─ YES → Google Vision API or AWS Textract
└─ NO → Local solution
    ├─ CJK (Chinese/Japanese/Korean) or tables? → PaddleOCR
    ├─ General photos or multiple languages? → EasyOCR
    ├─ Simple printed English docs? → Tesseract
    └─ PDF documents with structure? → Docling or Surya

Python Implementations

Tesseract (pytesseract)

import pytesseract
from PIL import Image
import cv2
import numpy as np

def extract_text_tesseract(image_path: str, lang: str = "eng") -> str:
    """Extract text using Tesseract. Best for clean printed documents."""
    image = Image.open(image_path)

    # Config: --psm 6 = assume uniform block of text
    config = "--psm 6 --oem 3"
    text = pytesseract.image_to_string(image, lang=lang, config=config)
    return text.strip()

def extract_with_confidence(image_path: str) -> list[dict]:
    """Extract text with bounding boxes and confidence scores."""
    image = Image.open(image_path)
    data = pytesseract.image_to_data(image, output_type=pytesseract.Output.DICT)

    results = []
    for i, word in enumerate(data["text"]):
        if word.strip() and int(data["conf"][i]) > 30:
            results.append({
                "text": word,
                "confidence": data["conf"][i],
                "bbox": {
                    "x": data["left"][i],
                    "y": data["top"][i],
                    "width": data["width"][i],
                    "height": data["height"][i],
                }
            })
    return results

# Install: pip install pytesseract pillow
# System: apt install tesseract-ocr (Linux) / brew install tesseract (Mac)

EasyOCR

import easyocr
from pathlib import Path

def extract_text_easyocr(
    image_path: str,
    languages: list[str] = ["en"],
    detail: bool = False
) -> str | list:
    """
    Extract text using EasyOCR. Best for photos and multiple languages.
    languages: ['en'], ['en', 'es'], ['ch_sim', 'en'], etc.
    """
    reader = easyocr.Reader(languages, gpu=False)  # gpu=True if CUDA available
    results = reader.readtext(image_path)

    if not detail:
        # Return plain text sorted by vertical position
        results_sorted = sorted(results, key=lambda x: x[0][0][1])
        return "\n".join([text for _, text, conf in results_sorted if conf > 0.3])

    return [
        {
            "text": text,
            "confidence": round(conf, 3),
            "bbox": bbox,
        }
        for bbox, text, conf in results
    ]

# Install: pip install easyocr

PaddleOCR (best for CJK and tables)

from paddleocr import PaddleOCR
import json

def extract_text_paddle(
    image_path: str,
    lang: str = "en",  # "en", "ch", "japan", "korean", "es", etc.
    use_angle_cls: bool = True,
) -> str:
    """Extract text using PaddleOCR. Best for CJK and structured documents."""
    ocr = PaddleOCR(use_angle_cls=use_angle_cls, lang=lang, show_log=False)
    result = ocr.ocr(image_path, cls=True)

    lines = []
    if result and result[0]:
        # Sort by y position (top to bottom)
        items = sorted(result[0], key=lambda x: x[0][0][1])
        lines = [item[1][0] for item in items if item[1][1] > 0.3]

    return "\n".join(lines)

# Install: pip install paddlepaddle paddleocr

Google Vision API

from google.cloud import vision
import io

def extract_text_google_vision(image_path: str) -> dict:
    """
    Extract text using Google Vision API.
    Requires: GOOGLE_APPLICATION_CREDENTIALS env var set.
    """
    client = vision.ImageAnnotatorClient()

    with io.open(image_path, "rb") as image_file:
        content = image_file.read()

    image = vision.Image(content=content)

    # Full text detection (better for documents)
    response = client.document_text_detection(image=image)
    document = response.full_text_annotation

    return {
        "text": document.text,
        "pages": [
            {
                "blocks": [
                    {
                        "text": " ".join(
                            symbol.text
                            for para in block.paragraphs
                            for word in para.words
                            for symbol in word.symbols
                        ),
                        "confidence": block.confidence,
                    }
                    for block in page.blocks
                ]
            }
            for page in document.pages
        ]
    }

# Install: pip install google-cloud-vision

AWS Textract (best for forms and invoices)

import boto3
import json

def extract_text_textract(image_path: str, region: str = "us-east-1") -> dict:
    """
    Extract text, forms, and tables using AWS Textract.
    Handles key-value pairs and structured tables automatically.
    """
    client = boto3.client("textract", region_name=region)

    with open(image_path, "rb") as f:
        image_bytes = f.read()

    response = client.analyze_document(
        Document={"Bytes": image_bytes},
        FeatureTypes=["TABLES", "FORMS"]
    )

    # Extract raw text
    blocks = response["Blocks"]
    lines = [b["Text"] for b in blocks if b["BlockType"] == "LINE"]

    # Extract key-value pairs (forms)
    key_values = {}
    key_map = {b["Id"]: b for b in blocks if b["BlockType"] == "KEY_VALUE_SET" and "KEY" in b.get("EntityTypes", [])}
    value_map = {b["Id"]: b for b in blocks if b["BlockType"] == "KEY_VALUE_SET" and "VALUE" in b.get("EntityTypes", [])}

    for key_block in key_map.values():
        key_text = _get_text_from_block(key_block, blocks)
        for rel in key_block.get("Relationships", []):
            if rel["Type"] == "VALUE":
                for val_id in rel["Ids"]:
                    if val_id in value_map:
                        val_text = _get_text_from_block(value_map[val_id], blocks)
                        key_values[key_text] = val_text

    return {
        "text": "\n".join(lines),
        "form_fields": key_values,
    }

def _get_text_from_block(block, all_blocks):
    word_ids = []
    for rel in block.get("Relationships", []):
        if rel["Type"] == "CHILD":
            word_ids.extend(rel["Ids"])

    block_map = {b["Id"]: b for b in all_blocks}
    words = [block_map[wid]["Text"] for wid in word_ids if wid in block_map and block_map[wid]["BlockType"] == "WORD"]
    return " ".join(words)

# Install: pip install boto3

Image Preprocessing

Preprocessing is the #1 factor in OCR accuracy. Always apply before running OCR.

import cv2
import numpy as np
from PIL import Image, ImageEnhance, ImageFilter

def preprocess_for_ocr(image_path: str, output_path: str = None) -> np.ndarray:
    """
    Full preprocessing pipeline for maximum OCR accuracy.
    Apply selectively based on image type.
    """
    img = cv2.imread(image_path)

    # 1. Convert to grayscale
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    # 2. Resize if too small (OCR works better at 300+ DPI)
    height, width = gray.shape
    if width < 1000:
        scale = 2000 / width
        gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)

    # 3. Deskew (fix rotation)
    gray = deskew(gray)

    # 4. Denoise
    denoised = cv2.fastNlMeansDenoising(gray, h=10)

    # 5. Binarization (choose one based on lighting)
    # Option A: Otsu (uniform lighting)
    _, binary = cv2.threshold(denoised, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

    # Option B: Adaptive (uneven lighting, shadows)
    # binary = cv2.adaptiveThreshold(denoised, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
    #                                 cv2.THRESH_BINARY, 11, 2)

    # 6. Morphological cleanup (remove noise dots)
    kernel = np.ones((1, 1), np.uint8)
    cleaned = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)

    if output_path:
        cv2.imwrite(output_path, cleaned)

    return cleaned

def deskew(image: np.ndarray) -> np.ndarray:
    """Correct image rotation using projection analysis."""
    coords = np.column_stack(np.where(image > 0))
    angle = cv2.minAreaRect(coords)[-1]

    if angle < -45:
        angle = -(90 + angle)
    else:
        angle = -angle

    if abs(angle) < 0.5:  # Skip if nearly straight
        return image

    h, w = image.shape
    center = (w // 2, h // 2)
    M = cv2.getRotationMatrix2D(center, angle, 1.0)
    return cv2.warpAffine(image, M, (w, h), flags=cv2.INTER_CUBIC,
                          borderMode=cv2.BORDER_REPLICATE)

def enhance_contrast(image_path: str) -> Image.Image:
    """Enhance contrast using PIL - useful for faded text."""
    img = Image.open(image_path).convert("L")
    enhancer = ImageEnhance.Contrast(img)
    return enhancer.enhance(2.0)

# Install: pip install opencv-python pillow

Preprocessing Decision Guide

PDF to Text Extraction

import fitz  # PyMuPDF - for native text extraction
from pdf2image import convert_from_path  # for scanned PDFs
import pytesseract

def extract_pdf_text(pdf_path: str, ocr_fallback: bool = True) -> str:
    """
    Smart PDF extraction:
    - Uses native text layer if available (fast, accurate)
    - Falls back to OCR for scanned pages
    """
    doc = fitz.open(pdf_path)
    full_text = []

    for page_num, page in enumerate(doc):
        # Try native text extraction first
        text = page.get_text().strip()

        if text and len(text) > 50:
            full_text.append(text)
        elif ocr_fallback:
            # Scanned page — render and OCR
            pix = page.get_pixmap(dpi=300)
            img_path = f"/tmp/page_{page_num}.png"
            pix.save(img_path)

            ocr_text = pytesseract.image_to_string(img_path)
            full_text.append(ocr_text)

    doc.close()
    return "\n\n".join(full_text)

# Install: pip install PyMuPDF pdf2image pytesseract
# System: apt install poppler-utils (for pdf2image on Linux)

Post-Processing Extracted Text

import re
from difflib import SequenceMatcher

def clean_ocr_text(text: str) -> str:
    """Standard cleanup for OCR output."""
    # Remove non-printable characters
    text = re.sub(r"[^\x20-\x7E\n\t]", "", text)

    # Normalize whitespace
    text = re.sub(r" +", " ", text)
    text = re.sub(r"\n{3,}", "\n\n", text)

    # Fix common OCR misreads
    corrections = {
        r"\b0(?=[a-zA-Z])": "O",    # 0 misread as O before letter
        r"(?<=[a-zA-Z])0\b": "O",    # O misread as 0 after letter
        r"\bl\b": "I",               # lowercase l misread as I (context-dependent)
        r"rn": "m",                  # rn → m (common serif font error)
    }
    for pattern, replacement in corrections.items():
        text = re.sub(pattern, replacement, text)

    return text.strip()

def extract_structured_data(text: str) -> dict:
    """Extract common structured fields from OCR text."""
    patterns = {
        "email": r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b",
        "phone": r"[\+]?[(]?[0-9]{3}[)]?[-\s\.]?[0-9]{3}[-\s\.]?[0-9]{4,6}",
        "date": r"\b\d{1,2}[/-]\d{1,2}[/-]\d{2,4}\b",
        "amount": r"\$\s?\d+(?:,\d{3})*(?:\.\d{2})?",
        "url": r"https?://[^\s]+",
    }

    return {
        field: re.findall(pattern, text)
        for field, pattern in patterns.items()
    }

def merge_multiline_words(text: str) -> str:
    """Fix hyphenated words split across lines (common in PDFs)."""
    return re.sub(r"(\w)-\n(\w)", r"\1\2", text)

Node.js / TypeScript

// Using Tesseract.js (pure JS, no native deps needed)
import Tesseract from "tesseract.js";

async function extractText(imagePath: string, lang = "eng"): Promise<string> {
  const { data } = await Tesseract.recognize(imagePath, lang, {
    logger: () => {}, // suppress progress logs
  });
  return data.text.trim();
}

// With confidence filtering
async function extractWithConfidence(imagePath: string) {
  const { data } = await Tesseract.recognize(imagePath, "eng");

  return data.words
    .filter((word) => word.confidence > 70)
    .map((word) => ({
      text: word.text,
      confidence: word.confidence,
      bbox: word.bbox,
    }));
}

// Install: npm install tesseract.js



// Using Google Vision API from Node.js
import vision from "@google-cloud/vision";

const client = new vision.ImageAnnotatorClient();

async function extractTextCloud(imagePath: string): Promise<string> {
  const [result] = await client.documentTextDetection(imagePath);
  return result.fullTextAnnotation?.text ?? "";
}

// Install: npm install @google-cloud/vision

Claude Vision API for OCR

Use Claude's vision capability when you need structured extraction + understanding:

import anthropic
import base64
from pathlib import Path

def extract_with_claude(image_path: str, instruction: str = None) -> str:
    """
    Use Claude to extract and structure text from an image.
    Best when you need semantic understanding, not just raw text.
    """
    client = anthropic.Anthropic()

    image_data = base64.standard_b64encode(Path(image_path).read_bytes()).decode()
    ext = Path(image_path).suffix.lower()
    media_types = {".jpg": "image/jpeg", ".jpeg": "image/jpeg", ".png": "image/png", ".webp": "image/webp"}
    media_type = media_types.get(ext, "image/jpeg")

    prompt = instruction or (
        "Extract ALL text from this image exactly as it appears. "
        "Preserve the original structure, line breaks, and formatting. "
        "Return only the extracted text, nothing else."
    )

    message = client.messages.create(
        model="claude-opus-4-6",
        max_tokens=4096,
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "image",
                        "source": {
                            "type": "base64",
                            "media_type": media_type,
                            "data": image_data,
                        },
                    },
                    {"type": "text", "text": prompt},
                ],
            }
        ],
    )

    return message.content[0].text

# Example: structured invoice extraction
def extract_invoice(image_path: str) -> dict:
    result = extract_with_claude(
        image_path,
        instruction="""Extract all data from this invoice and return as JSON:
{
  "invoice_number": "",
  "date": "",
  "vendor": {"name": "", "address": "", "email": ""},
  "items": [{"description": "", "quantity": 0, "unit_price": 0, "total": 0}],
  "subtotal": 0,
  "tax": 0,
  "total": 0
}
Return only valid JSON, no explanation."""
    )
    import json
    return json.loads(result)

When to Use Claude vs Traditional OCR

Accuracy Benchmarks by Image Type

*Requires additional language pack installation

Common Errors and Fixes

Tesseract returns garbage text

• Cause : Image too small or too noisy
• Fix : Upscale 2x, apply denoising and binarization

EasyOCR misses text in columns

• Cause : Default layout analysis fails on multi-column
• Fix : Crop each column separately and OCR individually

PaddleOCR slow on CPU

• Cause : Large model loaded
• Fix : Use use_gpu=True if available, or use_angle_cls=False for horizontal text

Bounding boxes don't align with text

• Cause : Image was rotated before OCR
• Fix : Apply deskew() in preprocessing

Cloud API returns empty for some regions

• Cause : Low contrast or very small text
• Fix : Preprocess image, increase DPI, crop region of interest

PDF text layer has wrong encoding

• Cause : Non-standard font embedding
• Fix : Use fitz.Page.get_text("rawdict") to inspect encoding, or skip to OCR fallback

Quick Start Templates

Minimal local OCR (Python)

pip install easyocr
python -c "import easyocr; r=easyocr.Reader(['en']); print('\n'.join([t for _,t,c in r.readtext('image.png') if c>0.3]))"

Minimal cloud OCR (Node.js)

npm install tesseract.js
node -e "const T=require('tesseract.js'); T.recognize('image.png','eng').then(r=>console.log(r.data.text))"

Batch processing pipeline

from pathlib import Path
import easyocr

reader = easyocr.Reader(["en"], gpu=False)

def batch_ocr(folder: str, output_folder: str) -> None:
    Path(output_folder).mkdir(exist_ok=True)
    images = list(Path(folder).glob("*.{png,jpg,jpeg,tiff,bmp}"))

    for img_path in images:
        results = reader.readtext(str(img_path))
        text = "\n".join(t for _, t, c in results if c > 0.3)

        out_path = Path(output_folder) / f"{img_path.stem}.txt"
        out_path.write_text(text, encoding="utf-8")
        print(f"✓ {img_path.name} → {out_path.name}")

    print(f"\nProcessed {len(images)} images.")

batch_ocr("./images", "./output")

Rules

• Select the OCR engine based on the document type and accuracy requirements before writing code: Tesseract for local/offline simple documents, EasyOCR for multilingual handwriting, cloud APIs (Google Vision, AWS Textract) for production accuracy on structured documents
• Image preprocessing (grayscale conversion, binarization, deskew) is required before Tesseract and EasyOCR for non-ideal inputs — skipping it causes significant accuracy degradation
• OCR output must always be treated as unvalidated text — apply post-processing (regex, string normalization) before using extracted values in business logic
• Never pass sensitive document images to cloud OCR APIs without confirming data privacy and compliance requirements with the project owner
• Confidence scores from the OCR engine must be checked; results below the project-defined threshold must be flagged for human review rather than accepted automatically

Weekly Installs

689

Repository

fearovex/claude-config

First Seen

Mar 4, 2026

Security Audits

Gen Agent Trust HubPassSocketPassSnykFail

Installed on

gemini-cli688

codex688

opencode688

amp687

github-copilot687

cursor687