tsegGAN: A Generative Adversarial Network for Segmenting Touching Nontext Components From Text Ones in Handwriting

Segmentation of a touching component to separate its constituent text and nontext parts is always a very crucial but challenging task toward developing a comprehensive document image processing (DIP) system. This is because, irrespective of document types, either printed or handwritten, the nontext parts need to be suppressed first before processing the text parts through an optical character recognition (OCR) system. Although a good number of attempts have been made to address this issue for printed documents, the same for regular handwritten document images is almost none. However, the appearance of touching components where a nontext part gets joined with a text part is a common issue in freestyle handwriting. To this end, in this work, we tailor-make a generative adversarial network (GAN)-based model with a suitable loss function that we name tsegGAN. We also prepare an in-house data set by collecting touching components from different real-world handwritten documents to evaluate our model. The performance comparison of our model with state-of-the-art GAN models shows that tsegGAN has outperformed the others with a significant margin.