EasyOCR

Warnings

• gotcha EasyOCR runs on PyTorch. For GPU acceleration, you MUST install a CUDA-enabled version of PyTorch separately. `pip install easyocr` will only install a CPU-compatible PyTorch if it's not already present.
  Fix: After installing easyocr, install PyTorch with CUDA from the official PyTorch website based on your CUDA version (e.g., `pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118`). Then, initialize `easyocr.Reader(..., gpu=True)`.
• gotcha The first initialization of `easyocr.Reader` for a new language will download large model files (several hundreds of MBs). This requires an internet connection and can cause a delay.
  Fix: Ensure an active internet connection on first use. For production, consider pre-downloading models or packaging them with your application (refer to EasyOCR documentation for model locations).
• gotcha Not all languages are supported by pre-trained models. Specifying an unsupported language or a non-existent model name will lead to an error or unexpected behavior.
  Fix: Consult the official EasyOCR GitHub repository for the list of supported languages. If your language is not supported or performs poorly, consider training a custom model.
• gotcha The `detect_network` argument was introduced in v1.6.0 to allow specifying alternative text detectors (e.g., 'dbnet18'). Prior to this, only the CRAFT detector was available. If you upgrade from an older version or need to ensure a specific detector, you might need to explicitly set this argument.
  Fix: To use DBNET, initialize `reader = easyocr.Reader(['en'], detect_network='dbnet18')`. If you need to ensure the CRAFT detector (the default prior to v1.6.0), you can explicitly set `detect_network='craft'`.
• gotcha EasyOCR can be memory-intensive, especially when processing high-resolution images or multiple languages simultaneously, consuming significant RAM or VRAM.
  Fix: For large images, consider pre-processing to resize or split them. For memory-constrained environments, run on CPU (`gpu=False`) and process images in batches or sequentially. Monitor resource usage and adjust configurations accordingly.

Install

• pip install easyocr Basic installation (CPU-only)
• pip install easyocr opencv-python-headless Recommended installation (for server environments)
• pip install easyocr torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 Installation with GPU support (Example for CUDA 11.8)

Imports

• Reader

  import easyocr
  reader = easyocr.Reader(['en'])

Quickstart

This quickstart code creates a simple dummy image with text, then initializes an EasyOCR Reader for English and French. The first time you run this for a new language, it will download the necessary language models. It then performs OCR on the image and prints the detected text along with its confidence score. For GPU acceleration, ensure you have correctly installed PyTorch with CUDA support and set `gpu=True`.

import easyocr
import os
from PIL import Image, ImageDraw

# Create a dummy image for demonstration purposes
dummy_image_path = 'easyocr_demo_image.png'
img = Image.new('RGB', (400, 100), color = (255, 255, 255))
d = ImageDraw.Draw(img)
d.text((10, 10), "Hello EasyOCR!\nThis is a test.", fill=(0,0,0))
img.save(dummy_image_path)

# Initialize Reader with desired languages.
# Models are downloaded on first run for each language.
# Set gpu=True if you have CUDA-enabled PyTorch installed, otherwise leave as False.
reader = easyocr.Reader(['en', 'fr'], gpu=False)

# Perform OCR on the image file
# result will be a list of tuples: (bounding_box, text, confidence)
result = reader.readtext(dummy_image_path)

print(f"OCR Results for '{dummy_image_path}':")
for (bbox, text, conf) in result:
    print(f"  Text: '{text}', Confidence: {conf:.2f}")

# Clean up the dummy image
os.remove(dummy_image_path)