Scholarlink

<p>Reconstruction of crosscut and ripped documents is challenging because of irregular fragmentation, loss of text and structural continuity, and exponential difficulty of reassembling them. This paper proposes a new Explainable AI (XAI)-powered machine learning framework that combines deep learning, graph optimization, and reinforcement learning for high-accuracy automated document reconstruction and interpretability. A Siamese Networks, Convolutional Neural Networks (CNNs), and Graph Neural Networks (GNNs) hybrid model is created for fragment matching and alignment, whereas the Hungarian Algorithm and Deep Q-Network (DQN) are used for optimizing reassembly. For the sake of transparency, explainability methods like SHAP, LIME, and Grad-CAM are integrated so that forensic validation of AI-driven decisions is made possible. The proposed method is compared on both real and synthetic data sets, having a reconstruction accuracy of 94.6%, which is substantially better than classical methods. This framework has critical applications in forensic science, intelligence operations, and historical document restoration, setting a new standard for interpretable AI-driven document reconstruction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel explainable AI approach for reconstructing crosscut and hand-torn documents using machine learning

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Mukktayakka Rajashekhar Desai,
  • Anil Kannur

摘要

Reconstruction of crosscut and ripped documents is challenging because of irregular fragmentation, loss of text and structural continuity, and exponential difficulty of reassembling them. This paper proposes a new Explainable AI (XAI)-powered machine learning framework that combines deep learning, graph optimization, and reinforcement learning for high-accuracy automated document reconstruction and interpretability. A Siamese Networks, Convolutional Neural Networks (CNNs), and Graph Neural Networks (GNNs) hybrid model is created for fragment matching and alignment, whereas the Hungarian Algorithm and Deep Q-Network (DQN) are used for optimizing reassembly. For the sake of transparency, explainability methods like SHAP, LIME, and Grad-CAM are integrated so that forensic validation of AI-driven decisions is made possible. The proposed method is compared on both real and synthetic data sets, having a reconstruction accuracy of 94.6%, which is substantially better than classical methods. This framework has critical applications in forensic science, intelligence operations, and historical document restoration, setting a new standard for interpretable AI-driven document reconstruction.