Degraded Document Binarisation

Abstract

In this study, I explored degraded document binarization by reviewing two recent model frameworks and implementing their models using PyTorch. The first model is based on cGANs, specifically the DE-GAN [41] framework, which enhances degraded documents by restoring their quality prior to binarization. The second model employs vision transformers [40], inspired by the DocBinFormer architecture, which uses an autoencoder in both the encoder and decoder for effective binarization. Both models were evaluated on the ISI-Bengali dataset. Experimental results demonstrate that DE-GAN improved document quality by 4% compared to the degraded input, while the vision transformer model achieved a 14% improvement, highlighting the effectiveness of transformer-based approaches for document enhancement and binarization.