Toward a Robust Segmentation Module Based on Deep Learning Approaches Resolving Historical Cursive Fonts Challenges

Historical documents are of major importance in conserving cultural and scientific heritage. In order to preserve this patrimony and to allow researchers from multiple fields to manipulate them, experts had opted to digitize these documents. In order to accelerate the process of classification while building logical connections between elements, the segmentation task should be treated carefully. Multiple non-textual blocks are presented within historical documents. Moreover, the non-consideration of the presence of columns generates major issues within the different processing stages. Based on deep learning segmentation approaches, this paper introduces the opted design for text/non-text separation and columns segmentations tasks. Moreover, a novel segmentation approach dealing with cursive fonts is proposed. The accuracy of the text/non-text blocks segmentation accuracy is equal to 97.32%. The columns segmentation approach accuracy is equal to 93.26%.