Deep Learning-based Reconstruction of Shredded Documents

Abstract

This thesis addresses the reconstruction of shredded paper documents, a relevant task in various domains such as forensic investigation and history reconstruction. Despite previous research, dealing with real-shredded data is a sensitive issue in the literature. To face this challenge, we proposed two deep learning approaches that have achieved state-of-the-art accuracy in more realistic scenarios. As a second major contribution, human interaction was explored to improve reconstruction. Our framework, inspired by the field of active learning, automatically selects potential mistakes in the solution for user analysis enabling better accuracy in a scalable way. The results yielded works in top-tier publications such as CVPR and the Pattern Recognition journal.