Word Segmentation in Urdu Text from News Video Tickers Using Deep Learning

Abstract

Automated word segmentation in Urdu text is a key problem in Optical Character Recognition (OCR) and Natural Language Processing (NLP) workflows because of its ligature-based Nastaliq script, in which Urdu is written from right to left. In broadcast news media, Urdu text emerges as a scrolling ticker integrated into video frames. Proper word segmentation is important for Urdu text recognition in video tickers. This paper introduces a deep learning-based method for segmentation of Urdu text tokens from news video tickers. We employed Fast Region-based Convolutional Neural Network (Fast R-CNN) with three transfer-learned feature extractors, a standard CNN, AlexNet, and VGG16, and train six detector configurations on the subsets of a custom annotated dataset, the Custom Urdu dataset (CUDS). CUDS comprises 2,515 scene images containing 27,199 labeled Urdu word instances collected from four Pakistani news television channels. Experimental results show that the best-performing model, F_R-CNN (CNN backbone, CUDS), achieves a precision of 1.0 and recall of 0.98 at 30 training epochs, while F_R-CNN (AlexNet, CUDS2) achieves precision and recall of 1.0. Comparative analysis across all six models demonstrates the strong suitability of Fast R-CNN for Urdu word segmentation in broadcast ticker footage.