CRNN-CTC Deep Learning Algorithm for Handwritten Amharic Document Recognition

Preserving handwritten Amharic documents for long periods of time requires converting them into a computer-readable form. In Ethiopia, Amharic is the official working language spoken by over 50 million people as a first language and over 100 million as a second language. As a result, there are a large number of handwritten documents in Amharic script. To address this, automatic transcription of handwritten Amharic documents is necessary. There have been a variety of attempts to develop Handwritten Character Recognition (HCR) systems for different language scripts, including Amharic, using both statistical and deep learning approaches. However, these systems have been hindered by the need for character segmentation and pre and Post processing, particularly in the Amharic script, which have a large number of characters, it is a challenging task. Recent advances in deep learning and access to advanced processors like GPU have made it possible to use sequence-to-sequence learning approaches for HCR. This thesis proposes a deep learning model that combines a Convolutional Neural Network (CNN) and a BidirectionalLSTM Network with Connectionist Temporal Classification (BiLSTM-CTC) to simplify the detailed processing needs of Amharic HCR. The model was trained using 1,800 line images and validated using 200 line images with dimensions of 32x256 and 48x256. The performance of the model was evaluated using the Character Error Rate (CER) on train, validation, and test data sets. The best trained model had a CER of 7.81% on test data with image dimensions of 48x256. These results demonstrate that the proposed sequence-to-sequence learning approach is a viable solution for HCR without the need for detailed image preprocessing and post-processing.