Text-line extraction from handwritten document images using GAN
作者：
Highlights：
• Applied Generative Adversarial Network first time for text line extraction.
• Superiority of U-Net architecture over encoder decoder in generator is shown.
• In discriminator PatchGAN is used to get the local pixel distribution.
• GAN loss, L1 loss and L2 loss are tested and merge suitably for better accuracy.
• It outperforms for HIT-MW and ICDAR2013 handwritten segmentation contest datasets.
摘要
•Applied Generative Adversarial Network first time for text line extraction.•Superiority of U-Net architecture over encoder decoder in generator is shown.•In discriminator PatchGAN is used to get the local pixel distribution.•GAN loss, L1 loss and L2 loss are tested and merge suitably for better accuracy.•It outperforms for HIT-MW and ICDAR2013 handwritten segmentation contest datasets.
论文关键词：GAN,Deep Learning,Text-line extraction,Handwritten documents,HIT-MW dataset,ICDAR dataset
论文评审过程：Received 17 March 2019, Revised 3 August 2019, Accepted 31 August 2019, Available online 2 September 2019, Version of Record 10 September 2019.
论文官网地址：https://doi.org/10.1016/j.eswa.2019.112916
