Text Line Detection and Recognition of Greek Polytonic Documents.

In this work we highlight the significance of Text Line Detection in documents. By utilizing a well-known Deep Neural Network and by proposing some simple but efficient modifications applied during training of such models, we can achieve very accurate results in different datasets of high diversity. Moreover, such models can be robust even when trained with few data. Our focus is on Greek polytonic documents (typewritten and handwritten) and we provide a new dataset to the public (GTLD-small) for text line detection. We evaluate our method through scenarios applied to the detection and recognition tasks, while demonstrating promising results when compared to popular commercial or open-source systems.