IndOCR-24: An Industrial OCR Dataset for Restricted-Alphabet Character Classification

Description

IndOCR-24 is a dataset developed for research on industrial Optical Character Recognition (OCR) involving restricted-alphabet character classification. The dataset was created from real industrial character images extracted from traceability codes printed on copper cathodes under operational conditions. The repository contains the complete set of original industrial character images, representative samples from the augmented training and validation datasets, and the Python scripts required to reproduce the complete augmented dataset used in the associated experiments. The augmentation pipeline applies photometric, geometric, and morphological transformations to the original images in order to increase variability while preserving the original class labels. The dataset is intended for benchmarking OCR algorithms, character classification methods, deep learning models, transfer learning, and data augmentation strategies in industrial computer vision applications.

Steps to reproduce

1. Download the repository and extract its contents. 2. Install the required Python packages, including Python 3, OpenCV, NumPy, and PyTorch. 3. Organize the original industrial character images into class-specific directories under the input dataset folder. 4. Execute the script `build_char_aug_dataset.py` to generate the augmented training and validation datasets. The script normalizes all images to 64 × 64 pixels and applies reproducible photometric, geometric, and morphological augmentation operations using a fixed random seed. 5. The generated datasets are automatically organized into `train` and `val` directories, and the corresponding annotation files (`train.csv` and `val.csv`) are created. 6. To reproduce the baseline classification experiments, execute the script `train_char_classifier.py` using the generated CSV files. The script trains a convolutional neural network and saves the best-performing model based on validation accuracy. 7. Representative samples of the augmented datasets are included in this repository. The complete augmented dataset used in the experiments can be regenerated by executing the provided scripts.

Institutions

• Austral University of Chile

  Los Ríos Region, Valdivia

Categories

Licence