TRANSFER LEARNING FOR PRINTED PASHTO ALPHABET RECOGNITION: A COMPARATIVE ANALYSIS

Abstract

Optical character recognition (OCR) has made significant improvements thanks to deep learning, but scripts with few resources, like Pashto, are still not well covered in research. The Pashto script is written from right to left and has a cursive style. Also, some characters look very similar but are different because of small dots or marks. This makes it hard to recognize characters accurately. Most past studies have used custom-built convolutional neural network (CNN) models for recognizing Pashto script, but very few have compared common pretrained CNN models. This paper fills that gap by testing three pretrained CNN models—MobileNetV2, EfficientNetB0, and ResNet50—for recognizing printed Pashto letters across 44 different characters. All models were trained and tested under the same conditions, and their performance was measured using test accuracy, macro precision, macro recall, and macro F1-score. The results showed that ResNet50 performed the best, with a test accuracy of 87.39% and macro precision, recall, and F1-scores of 0.89, 0.87, and 0.87, respectively. This is better than MobileNetV2 (82.30%) and EfficientNetB0 (76.40%). ResNet50 also showed better generalization, performing best in both validation and test phases even though its training accuracy was lower than MobileNetV2. This suggests that ResNet50 learns more robust features for distinguishing similar Pashto characters. The study shows that choosing the right pretrained CNN model greatly affects recognition results for scripts with limited resources. It also shows that ResNet50 is the most effective model among those tested. This work sets up a clear benchmark for using transfer learning in printed Pashto character recognition and adds to the limited research on Pashto OCR.