Character recognition of ancient ethiopic Ge'ez manuscripts using deep convolutional neural networks

The Ethiopian Orthodox Tewahedo Church, along with the Eritrean Orthodox Tewahedo Church, uses the ancient language Ge'ez for its liturgy. This preconial language, rich in history, plays a central role in religious texts. With the rise of technological advancements, particularly in optical character recognition (OCR), the need to digitize Ge'ez manuscripts and make them accessible to the public has become increasingly important. In an OCR system, the text is digitized, processed, and segmented, and features are extracted, categorized, and recognized. The first step in character recognition involves digitizing documents via tools such as the Cam Scanner at a resolution of 300 dpi. Once scanned, the image documents are preprocessed to remove noise via a set of low-pass filters. In our study, four low-pass algorithms for noise reduction and four thresholding algorithms for image binarization were tested. On the basis of having the lowest mean square error and highest peak signal-to-noise ratio, we selected bilateral filtering for noise reduction and Otsu’s thresholding for binarization. After preprocessing, the next step is image segmentation, where lines and individual characters are detected via a bounding box or bounding rectangle algorithm. Since the segmented characters vary in shape due to handwriting differences, they are normalized to a uniform size of 32 × 32 pixels. Character features are then extracted via automatic feature extraction methods within a deep learning framework. The recognition model is developed via pretrained deep convolutional neural networks, such as ResNet50 and VGG19, along with CNN-based support vector machines, trained on collected data. The experimental results show that a sequentially developed deep convolutional neural network achieves the highest accuracy of 93.52% in character recognition. However, segmentation errors, often caused by the degradation of document images, negatively impact the performance of the recognition models. Future research should focus on implementing advanced image enhancement and restoration techniques to further improve accuracy.