Typewritten OCR Model for Ethiopic Characters

Abstract

Optical Character Recognition (OCR) is the electronic conversion of images of computer-written, typewritten, handwritten, or printed text into machine-encoded text from a scanned document and a photo of a document. In Ethiopia, documents such as historical, office, and official documents have been documented in handwritten and typewritten form, until recently. Thus, a large number of historical and essential documents are still in hard-copy form and at risk of disaster to be lost. Computer-written and handwritten OCR have been developed for different language characters including Ethiopian languages (Ethiopic characters), But not typewriter-written OCR for Ethiopic scripts. Like handwritten documents, large historical documents are typewritten documents in Ethiopia. Thus, the typewritten OCR is mandatory to preserve these documents. This study focuses on building an OCR model for typewritten documents that are written on Ethiopic characters. For the study, different Ethiopic characters have been collected from typewritten documents, and 290 distinct characters have been segmented to construct augmented data to form various character variations and simulate the complexities encountered in real-world typewritten Amharic texts and enhance the adaptability of the OCR model. This technique aims to approximate the diversity inherent in the data. The model training framework leverages the capabilities of Tesseract, an open-source OCR engine, in conjunction with the artificially generated training set. The Tesseract’s existing Amharic OCR model has been deployed as a base model, and the fine-tuning process has been adopted in a layered approach by employing 45,000 samples and spanning 4,800 iterations. The model has been evaluated using character error rate (CER). As per the evaluation, the model performed with 13% CER on the test set. For this study, the Tesseract model before fine-tuning and the Google Lense platform has been used as a baseline to evaluate the performance of the model. Accordingly, our model has outperformed both baselines by more than 10% margin.