Dual-Stage Inpainting Approach for Character Reconstruction in Ancient Hindi Texts

The study of historical documents often encounters challenges in dealing with missing or deteriorated content, hindering comprehension and preservation efforts. This paper focuses on a foundational approach aimed at restoring characters in deteriorated documents written in the Hindi language. The primary objective of this paper is to address the issue of missing regions within individual Hindi characters prevalent in ancient texts. The solution involves a dual-stage approach integrating a custom EfficientNet model and a Deep Image Prior (DIP) model to paint the missing regions in the input images. Initially, the custom EfficientNet model serves as a feature extractor, identifying characteristics from the damaged characters in the input image. The DIP model starts with generating random noise, gradually evolving towards the predicted character class over multiple iterations following the weights provided by the EfficientNet model. The DIP approach shows promising results over various test inputs. The approach also remains competitive, with different levels of noise introduced to the input image. This approach keeps aside contextual complexities of the input to prioritize the restoration of individual characters within ancient texts, contributing to the broader goal of historical document preservation and comprehension.