Bilingual Documents Text Lines Extraction Using Conditional GANs

Abstract

Text line extraction in document recognition is the major step. A number of classical approaches are available like projection profile, bounding box analysis, etc. These classical approaches are unable to segment the text with large variations in individual handwriting. Furthermore, segmentation of documents having data from multiple scripts creates more hurdles due to the presence of different writing styles. The usage of deep networks has been less explored in this domain due to the need of high training time and data. In this research, we have used conditional generative adversarial networks (GANs) for text line extraction in bilingual documents containing Gurumukhi-Latin scripts. It considers text line segmentation problem as image-to-image translation task. Two kinds of encoder–decoder networks are used for comparison, i.e., with skip connections and without skip connections. Dataset for bilingual handwritten documents containing 150 document images has been designed. It includes large variability in writing style and content. Results on the designed dataset for text line extraction are efficient for encoder–decoder network with skip connections.