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

Self-supervised deep reconstruction of mixed strip-shredded text documents

“…Additionally, for better accuracy, the number and size of candidate sequences have to be increased, which compromises the run-time performance (it performed approx. 16 times slower than [15] for a 60-shred instance).…”

“…In the preliminary work [15], the experiments were not cross-database since the documents of S-Isri-OCR were reconstructed with a model trained on documents of the ISRI-OCR Tk collection. In practice, such experiments assume the availability of training data that share significant appearance and structural similarities with test data.…”

“…A central procedure in their method is sampling candidate sequences and applying OCR for word recognition to improve pair compatibility estimation. Despite reporting accuracy comparable to [15] (the preliminary investigation on deep learning-based reconstruction), their validation relies on solely three real-shredded instances. For two of them, those which are up to 39 shreds, their method achieved accuracy above 70%, while the third instance, with 56 shreds, yielded 41.8% of accuracy.…”

“…Other specific contributions of this work include: (i) a more in-depth description of the technical details; (ii) an extended discussion of the state-of-the-art; (iii) an ablation study to analyze the sensitivity of the method to some key parameters. Experiments were conducted on the two collections of shredded documents used in [15], as well as on the newly-introduced dataset. Results have shown that our deep learning approach outperformed the competing methods, being capable of reconstructing 100 mixed shredded documents (2,292 shreds) with accuracy superior to 90%, which brings the state-of-theart of the document reconstruction problem to another level.…”

“…This enables the self-supervised learning of the model, in which class labels are deter-mined directly from adjacency information of the shreds. A preliminary investigation of this approach [15] has indicated the potential of using simulated data to reconstruct real-shredded documents in constrained scenarios, such as single-page reconstruction and intra-database evaluation. The first scenario assumes that the shreds are previously separated by pages so that the reconstruction system should deal with one page at a time (limited to 30 shreds per page).…”