High Performance Urdu and Arabic Video Text Recognition Using Convolutional Recurrent Neural Networks

Text extraction from videos is an emerging research field in the document analysis community. We propose a simple Convolutional Recurrent Neural Network to perform text recognition on both Arabic and Urdu scripts. We use a large variety of data augmentation techniques to generalize the model and prevent over-fitting. We also use a slightly improved loss function that helps the model converge faster. Using the proposed method we achieved 99.73% CRR, 88.37% WRR and 89.92% LRR on the Urdu Ticker Text dataset and 96.82% CRR, 90.41% WRR and 76.78% LRR on the AcTiVComp20 dataset. The proposed method has significantly outperformed Google Vision API on both of the datasets.