PHND: Pashtu Handwritten Numerals Database and deep learning benchmark.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Pashtu is spoken by over 50 million people, yet lacks dedicated Optical Character Recognition (POCR) resources.
• Existing research has not significantly addressed Pashtu handwritten numeral recognition (PHNR).

Purpose of the Study:

• To introduce the first large-scale Pashtu Handwritten Numerals Dataset (PHND) with 50,000 images.
• To present a novel deep neural network-based approach for PHNR.

Main Methods:

• Development of a new dataset (PHND) for Pashtu handwritten numerals.
• Implementation of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for feature extraction and classification.
• Training and evaluation of deep learning models on the PHND and CMATERDB 3.1.1 datasets.

Main Results:

• Achieved a top recognition rate of 98.00% on the Pashtu Handwritten Numerals Dataset (PHND).
• Obtained a recognition rate of 98.64% on the Bangla CMATERDB 3.1.1 dataset, demonstrating model generalizability.
• Successfully demonstrated the efficacy of deep neural networks for PHNR.

Conclusions:

• The proposed deep learning approach significantly advances Pashtu handwritten numeral recognition.
• The publicly available PHND dataset will facilitate future research in POCR.
• This study addresses a critical gap in low-resource language OCR technology.