Reconstruction of Shredded Figures Based on Grayscale Matrix

With the rapid advancement of information technology, image processing methodologies have become increasingly central to forensic analysis, historical document restoration, and intelligence extraction. In forensic applications, such methodologies facilitate the examination of evidentiary materials; in cultural heritage contexts, they support the recovery of damaged documents. Manual reconstruction of shredded documents is laborious and time-consuming, while existing automated approaches still face challenges in edge matching, semantic consistency preservation, and multi-modal data fusion. This paper presents a critical review of the strengths and limitations of current shredded document reconstruction algorithms and proposes tailored methods for mechanically shredded pages. In simulated experiments, edge features are extracted from real images through clustering; a combined metric — integrating a sum-of-differences (SODs) measure with correlation coefficient-based matching — is employed to evaluate boundary similarity and enable both horizontal and vertical reassembly of mechanically strip-cut fragments. Experimental results demonstrate that for single-direction (vertical) strips containing either Chinese or English text, the combined application of binarization, grayscale matrix edge analysis, SODs matching, and correlation matching enables reliable left-to-right recursive reconstruction, achieving complete restoration on the evaluated dataset. For cross-cut fragments, K-means clustering of row features combined with an iterative correlation threshold procedure reduces the need for manual intervention by over 40% while maintaining high reconstruction accuracy. For double-sided prints, comparing front-and-back correlation coefficients allows for accurate two-sided reconstruction. Furthermore, embedding edge features into a shared representation using a deep metric learning model trained with a contrastive loss function elevates single-page reconstruction accuracy to over 90%. Notable limitations remain: the limited processing capacity of typical hardware hinders the accurate modeling of gray-level transitions at authentic fragment joins. Future research will focus on simulating transition gray-value variations, incorporating edge irregularities (e.g. bevels and torn edges), and extending the metric evaluation to multi-page stitching, thereby enhancing the robustness and generalizability of the proposed approach.