Autonomous reconstruction of strip-shredded documents via self-supervised deep learning and global optimization

Abstract

Autonomous reconstruction of mechanically shredded documents is a labor-intensive challenge in forensic and archival workflows, particularly for scripts with complex structures such as Simplified Chinese. While traditional manual reassembly is tedious, existing digital tools typically rely on extensive human intervention. This paper presents an automated reassembly framework that integrates a lightweight convolutional feature extractor with global combinatorial optimization. By adapting the established SqueezeNet v1.1 backbone, we employ a task-specific self-supervised learning strategy trained on synthetically shredded samples, enabling the adapted model to capture local stroke continuity and edge-geometry cues without manual annotation. The framework infers pairwise relationships from calibrated edge-region inputs, organizing compatibility scores into an asymmetric traveling salesman problem (ATSP) formulation. The optimal fragment sequence is solved deterministically using the Concorde TSP solver, yielding a globally consistent reconstruction. Experimental results on physically shredded documents demonstrate reconstruction accuracies of 86.5% for Simplified Chinese and 94.8% for Western scripts. These results indicate that the proposed pipeline effectively generalizes from synthetic training data to real-world scenarios, providing a practical, high-throughput foundation for automated document recovery under computational constraints typical of robotic or embedded systems.