mazafard/portugues_ocr_dataset

📘 Portuguese OCR Dataset

This dataset contains scanned image-text pairs in European Portuguese, curated manually for Optical Character Recognition (OCR) tasks. It includes literary sentences and historical excerpts.

📦 Dataset Overview

• Total Samples: 10,000
• Image Shape: (10000, 100, 1200, 3) — color images with height 100 and width 1200 pixels
• Text Example:
  "E mais avante o Estreito que se arreia"

📂 Dataset Structure

The dataset is stored in a single HDF5 file (dataset.h5), which includes:

• images: a NumPy array of grayscale image data (from PNG files)
• texts: a list of UTF-8 encoded strings, corresponding to each image

Each sample pairs an image and its corresponding transcription.

📊 Dataset Statistics

• Format: HDF5 (dataset.h5)
• Total samples: ~10,000
• Average image size: 224x224 pixels (if preprocessed)
• Language: European Portuguese
• Source: literary and historic texts

💾 How to Load

To use this dataset with the Hugging Face Datasets library:

from datasets import load_dataset
import h5py

# Load from Hugging Face
dataset_path = "mazafard/portugues_ocr_dataset"
h5_file = load_dataset(dataset_path, data_files="dataset.h5", split="train")

# Alternatively, open the HDF5 file directly
with h5py.File("dataset.h5", "r") as f:
    images = f["images"][:]        # NumPy array of images
    texts = f["texts"][:]          # List of transcriptions

🧠 Use Cases

• Fine-tuning OCR models like microsoft/trocr-base-printed
• Document digitization
• Research in Portuguese language modeling and handwritten/printed recognition

📜 License

MIT License — free for academic and commercial use with attribution.

📦 How to Inspect dataset.h5

You can extract metadata from the HDF5 file using h5py:

import h5py

with h5py.File("dataset.h5", "r") as f:
    print("Keys:", list(f.keys()))  # ['images', 'texts']
    print("Number of samples:", len(f["texts"]))
    print("Image shape:", f["images"].shape)
    print("Example text:", f["texts"][0])

⚠️ Note

This dataset uses synthetic text-image pairs for printed OCR. Handwritten or scanned documents are not included.