Deep Learning Based Pashto Characters Recognition

In artificial intelligence, text identification and analysis that are based on images play a vital role in the text retrieving process. Automatic text recognition system development is a difficult task in machine learning, but in the case of cursive languages, it poses a big challenge to the research community due to slight changes in character’s shapes and the unavailability of a standard dataset. While this recognition task becomes more challenging in the case of Pashto language due to a large number of characters in its dataset than other similar cursive languages (Persian, Urdu, Arabic) and a slight change in character’s shape. This paper aims to address accept these challenges by developing an optimal optical character recognition (OCR) system to recognise isolated handwritten Pashto characters. The proposed OCR system is developed using multiple long short-term memory (LSTM) based deep learning model. The applicability of the proposed model is validated by using the decision trees (DT) classification tool based on the zoning feature extraction technique and the invariant moment approaches. An overall accuracy rate of 89.03% is calculated for the multiple LSTM-based OCR system while DT-based recognition rate of 72.9% is achieved using zoning feature vector and 74.56% is achieved for invariant moments-based feature map. Applicability of the system is evaluated using different performance metrics of accuracy, f-score, specificity, and varying training and test sets.