Ligature Recognition in Urdu Caption Text using Deep Convolutional Neural Networks

Textual content in videos contain rich information that can be exploited for semantic indexing and subsequent retrieval as well as development of video analytics solutions. The key modules in a textual content based video retrieval system include detection (localization) of text followed by its recognition, the later being the subject of our study. More specifically, this paper presents a caption text recognition system targeting Urdu text. The technique relies on a holistic approach using ligatures as units of recognition. Data driven feature extraction techniques are employed using a number of pre-trained deep convolution neural networks. The networks are used as feature extractors as well as fine-tuned on the ligature dataset under study and realized high ligature recognition rates.