A Deep Learning-Based Compatibility Score for Reconstruction of Strip-Shredded Text Documents

Abstract

The use of paper-shredder machines (mechanical shredding) to destroy documents can be illicitly motivated when the purpose is hiding evidence of fraud and other sorts of crimes. Therefore, reconstructing such documents is of great value for forensic investigation, but it is admittedly a stressful and time-consuming task for humans. To address this challenge, several computational techniques have been proposed in literature, particularly for documents with text-based content. In this context, a critical challenge for automated reconstruction is to measure properly the fitting (compatibility) between paper shreds (strips), which has been observed to be the main limitation of literature on this topic. The main contribution of this paper is a deep learning-based compatibility score to be applied in the reconstruction of strip-shredded text documents. Since there is no abundance of real-shredded data, we propose a training scheme based on digital simulated-shredding of documents from a well-known OCR database. The proposed score was coupled to a black-box optimization tool, and the resulting system achieved an average accuracy of 94.58% in the reconstruction of mechanically-shredded documents.