Shan_OCR_Dataset

This dataset was developed to support research on optical character recognition (OCR) for the Shan (Tai Long) script. It contains line-level Shan text data and corresponding UTF-8-encoded ground-truth transcriptions.The text-line images were preprocessed for sequence-based text recognition using a standardized input height of 54 pixels and variable widths to preserve the original aspect ratios and differences in text-line length. Following preprocessing, the images were converted into NumPy array files in .npy format. Consequently, the data can be loaded directly into the model-training and evaluation pipelines without requiring additional image preprocessing. This preparation reduces processing time and ensures consistent input representations across experiments.The dataset was used to train and evaluate Convolutional Recurrent Neural Network (CRNN) models incorporating Bidirectional Long Short-Term Memory (BiLSTM) layers and Connectionist Temporal Classification (CTC).The repository contains one training dataset and two evaluation datasets. The training data and their corresponding labels are provided in Train_data (1).zip and Train_label.txt. The evaluation data are divided into a Shan-only dataset, provided in Shan_Dataset.zip with its corresponding Shan_labels.txt, and a mixed-language dataset, provided in Mixed_test_dataset.zip with its corresponding Mixed_Label.txt. The mixed-language dataset is intended to evaluate the model’s recognition capability when Shan text occurs together with additional scripts or languages represented in the dataset.The dataset is intended to facilitate reproducible research on Shan-script OCR, particularly the recognition of complex characters, vowels, diacritics, combining marks, and stacked glyph structures. It may also support comparative evaluation and further development of OCR architectures for Shan and other low-resource writing systems.The repository includes preprocessed .npy data, corresponding ground-truth label files, and dataset documentation required to understand and reproduce the experiments described in the associated IEEE Access article.