Ethiopic natural scene text recognition using deep learning approaches

Abstract

The success of deep learning approaches for scene text recognition in English, Chinese and Arabic language inspired us to pose a benchmark scene text recognition for Ethiopic script. To transcribe the word images to the cross bonding text, we use a segmentation free end-to-end trainable Convolutional and Recurrent Neural Network (CRNN) hybrid architecture. In the network, robust representation features from cropped word images are extracted at convolutional layer and the extracted representations features are transcribed to a sequence of labels by the recurrent layer and transcription layer. The transcription is not bounded by lexicon or word length. Due to it is effective uses to transcribe sequence-to-sequence tasks, CTC loss is applied to train the network. In order to train the proposed model, we prepare synthetic word images from Unicode fonts of Ethiopic scripts, besides the model performance is evaluated on real scene text dataset collected from different sources. The experiment result of the proposed model, shows a promising result.