BetterOCR

🔍 Better text detection by combining multiple OCR engines with 🧠 LLM.

OCR still sucks! ... Especially when you're from the other side of the world (and face a significant lack of training data in your language) — or just not thrilled with noisy results.

BetterOCR combines results from multiple OCR engines with an LLM to correct & reconstruct the output.

🔍 OCR Engines

Currently supports EasyOCR (JaidedAI), Tesseract (Google), and Pororo (KakaoBrain).

• For Pororo, we're using the code from https://github.com/black7375/korean_ocr_using_pororo
  (Pre-processing ➡️ Text detection with EasyOCR ➡️ Text recognition with BrainOCR (Pororo's OCR module)).
• Pororo is used only if the language options (lang) specified include either 🇺🇸 English (en) or 🇰🇷 Korean (ko). Also additional dependencies listed in [tool.poetry.group.pororo.dependencies] must be available. (If not, it'll automatically be excluded from enabled engines.)

🧠 LLM

Supports Chat models from OpenAI.

📒 Custom Context

Allows users to provide an optional context to use specific keywords such as proper nouns and product names. This assists in spelling correction and noise identification, ensuring accuracy even with rare or unconventional words.

🛢️ Resources

• Head over to 💯 Examples to view performace by languages (🇺🇸, 🇰🇷, 🇮🇳).
• Coming Soon: box detection 🧪✅, improved interface 🚧, async support, and more. Contributions are welcomed.

Warning
This package is under rapid development 🛠

Architecture

🚀 Usage (WIP)

pip install betterocr
# pip3 install betterocr

import betterocr

# text detection
text = betterocr.detect_text(
    "demo.png",
    ["ko", "en"], # language codes (from EasyOCR)
    context="", # (optional) context
    tesseract={
      # Tesseract options here
      "config": "--tessdata-dir ./tessdata"
    },
    openai={
      # OpenAI options here

      # `os.environ["OPENAI_API_KEY"]` is used by default
      "API_KEY": "sk-xxxxxxx",

      # rest are used to pass params to `client.chat.completions.create`
      # `{"model": "gpt-4"}` by default
      "model": "gpt-3.5-turbo",
    },
)
print(text)

📦 Box Detection

Original	Detected

Example Script: https://github.com/junhoyeo/BetterOCR/blob/main/examples/detect_boxes.py (Uses OpenCV and Matplotlib to draw rectangles)

import betterocr

image_path = ".github/images/demo-1.png"
items = betterocr.detect_boxes(
    image_path,
    ["ko", "en"],
    context="퍼멘테이션 펩타인 아이케어 크림",  # product name
    tesseract={
        "config": "--psm 6 --tessdata-dir ./tessdata -c tessedit_create_boxfile=1"
    },
)
print(items)

[
  {'text': 'JUST FOR YOU', 'box': [[543, 87], [1013, 87], [1013, 151], [543, 151]]},
  {'text': '이런 분들께 추천드리는 퍼멘테이션 펩타인 아이케어 크림', 'box': [[240, 171], [1309, 171], [1309, 224], [240, 224]]},
  {'text': '매일매일 진해지는 다크서클을 개선하고 싶다면', 'box': [[123, 345], [1166, 345], [1166, 396], [123, 396]]},
  {'text': '축축 처지는 피부를 탄력 있게 바꾸고 싶다면', 'box': [[125, 409], [1242, 409], [1242, 470], [125, 470]]},
  {'text': '나날이 늘어가는 눈가 주름을 완화하고 싶다면', 'box': [[123, 479], [1112, 479], [1112, 553], [123, 553]]},
  {'text': 'FERMENATION', 'box': [[1216, 578], [1326, 578], [1326, 588], [1216, 588]]},
  {'text': '민감성 피부에도 사용할 수 있는 아이크림을 찾는다면', 'box': [[134, 534], [1071, 534], [1071, 618], [134, 618]]},
  {'text': '얇고 예민한 눈가 주변 피부를 관리하고 싶다면', 'box': [[173, 634], [1098, 634], [1098, 690], [173, 690]]}
]

💯 Examples

Note
Results may vary due to inherent variability and potential future updates to OCR engines or the OpenAI API.

Example 1 (English with Noise)

Example 2 (Korean+English)

Example 3 (Korean with custom context)

🧠 LLM Reasoning (*Old)

Based on the given OCR results and the context, here is the combined and corrected result:

{
  "data": "바이오힐보\n#세로모공쫀쫀세럼\n글로우픽 설문단 100인이 꼼꼼하게 평가했어요\n#누적 판매액 40억\n#제품만족도 100%"
}

• 바이오힐보 is the correct brand name, taken from [1] and the context.
• #세로모공쫀쫀세럼 seems to be the product name and is derived from the context.
• 글로우픽 설문단 100인이 꼼꼼하게 평가했어요 is extracted and corrected from both OCR results.
• #누적 판매액 40억 is taken from [0].
• #제품만족도 100% is corrected from [0].

Example 4 (Hindi)

License (Starmie!)

MIT © Junho Yeo

If you find this project interesting, please consider giving it a star(⭐) and following me on GitHub. I code 24/7 and ship mind-breaking things on a regular basis, so your support definitely won't be in vain!