Semi-Automated Transcription Generation for Pashto Cursive Script

Usually, a large amount of transcription data is required for training and benchmarking Optical Character Recognition (OCR) systems for new scripts like Pashto. In case of real image data; mostly the images are acquired through scanning. For supervised training scenarios, it is required to have a ground truth against the corresponding scanned images. Usually, the ground truth is created by transcribing the documents manually, which is an overwhelmingly laborious phase. This work introduces a semi-automated procedure for transcribing Pashto document images using a long short term memory (LSTM) network architecture. The process is applied for the transcription of 1000 images having Pashto ligatures and it improves the transcription performance to around three times of manual method.