In this paper, we introduce an end-to-end Amharic text-line image recognition approach based on recurrent neural networks. Amharic is an indigenous Ethiopic script which follows a unique syllabic writing system adopted from an ancient Geez script. This script uses 34 consonant characters with the seven vowel variants of each (called basic characters) and other labialized characters derived by adding diacritical marks and/or removing parts of the basic characters. These associated diacritics on basic characters are relatively smaller in size, visually similar, and challenging to distinguish from the derived characters. Motivated by the recent success of end-to-end learning in pattern recognition, we propose a model which integrates a feature extractor, sequence learner, and transcriber in a unified module and then trained in an end-to-end fashion. The experimental results, on a printed and synthetic benchmark Amharic Optical Character Recognition (OCR) database called ADOCR, demonstrated that the proposed model outperforms state-of-the-art methods by 6.98% and 1.05%, respectively.

With the increasing digitization of printed materials, it has become common to encounter documents in multiple languages in our daily lives. This has led to significant interest from researchers in the field of document image analysis, which includes applications like OCR, document image retrieval, and machine translation. However, these applications typically assume that documents are written in a single language, which is not always the case. Analyzing multilingual documents can be challenging, especially when the languages are switched within the document and the scripts are the same. In particular, this issue for low-resource languages has not yet been addressed. Hence, this paper focuses on exploring the use of deep learning techniques to develop a language identification model for multilingual documents written in Ethiopic script. To train and test our model, we extracted a total of 24,444 and 32,000 text-line pictures from a set of 814 and 1066 documents written in Ethiopian script and obtained from a variety of sources, respectively. We also carried out extensive experiments using both pre-trained CNN models and models trained from scratch with different parameter tuning on printed and synthetic document image datasets. The results revealed that CNN outperformed pre-trained deep learning models, achieving a promising accuracy of 97.27% on a test set of 9,600 images. These findings show the potential for developing effective language identification solutions for document images written in Ethiopic script.