Segmentation and Recognition of Text Images Acquired by a Mobile Phone

Segmentation and recognition of text in document images are two important steps in a document image understanding system. Several systems are proposed and used to ensure these steps, but less attention has been given to the images that are obtained by a mobile terminal. In order to overcome this limitation, we present in this paper a new text printed recognition system of document images obtained via a smartphone. Firstly, we apply a pre-processing step to extract and enhance the text region, after that we propose a new text-line segmentation algorithm that based on connected components (CCs) analysis in order to segment the text in individual lines. Finally, a bidirectional recurrent neural network (BRNN) with Gated Recurrent Unit (GRU) is trained to recognize the text-lines image. We evaluated the proposed system on ICDAR2015 Smartphone document OCR dataset. Experimental results demonstrate that BRNN-GRU model performs better with a higher computational speed compared to Long Short Time Memory (LSTM) that often used in the text recognition system.