Pashto Language Handwritten Numeral Classification Using Convolutional Neural Networks

Efficiency of the prevailing algorithms for recognizing hand-written text is constrained by the suboptimal performance of character recognition techniques applied to such images. Intricate backgrounds, diverse writing styles, varying text sizes and orientations, low resolutions, and presence of multi-language text collectively render the task of text recognition in natural images extremely complex and challenging. While conventional machine learning approaches have demonstrated satisfactory outcomes, the recognition of cursive text like Arabic, Urdu, and Pashto scripts in natural images remains an ongoing research challenge. Recognizing handwritten text poses a significant challenge when it comes to accurately segmenting and identifying individual characters. Variations in character shapes caused by their positions within words further compound the complexity of the recognition task. Optical character recognition (OCR) methods designed for Arabic, Urdu, and Pashto scanned documents show limited effectiveness when applied to character recognition in natural images. Keeping in view all these challenges we proposed a text classifier for Pashto handwritten digits based on a deep learning algorithm. Our proposed model achieves a classification accuracy of 99.05% on a publicly available Pashto Language Digit Dataset.