Bridging the Gap: A Novel Dataset for Handwriting-Aware Image Inpainting

DESCRIPTION

Image inpainting refers to the task in automated vision that aims to effectively reconstruct or fill in the missing, corrupted, or unnecessary areas of an image. Recent technological developments have enabled the seamless modification of digital or natural content. Unfortunately, these tools have introduced new challenges and risks of potential misuse and unintended damage to natural scenes and images through the inclusion of inappropriate text and handwritten overlays. The task of inpainting these texts has turned out to be challenging. Deep learning-based methods require a large amount of training data to perform appropriately. More than publicly available datasets are needed to address such issues. Observing this gap, we propose to build a handwritten image inpainting dataset. The proposed dataset provides a realistic benchmark for tasks that require to reconstruct natural images with complex structures like handwritten text, thereby facilitating the evolution of more robust models.

This dataset is provided exclusively for research purposes, and users are expected to cite following paper and any subsequent related papers that build upon this work.

[1] Somanka Maiti, Shabari Nath P and Gaurav Bhatnagar, Bridging the Gap: A Novel Dataset for Handwriting-Aware Image Inpainting, International Journal of Image, Graphics and Signal Processing, 2024. (Submitted)

DOWNLOAD LINK AND SAMPLE IMAGES

Twenty-five representative images of dataset are shown in the following figure. For complete dataset, please write an email to goravb@iitj.ac.in with the email subject "Requesting access to Handwriting-Aware Image Inpainting Dataset".