Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LayeredDoc: Domain Adaptive Document Restoration with a Layer Separation Approach

Maria Pilligua, Nil Biescas, Javier Vazquez-Corral, Josep Lladós, Ernest Valveny, Sanket Biswas

TL;DR

A text-graphic layer separation approach that enhances domain adaptability in document image restoration (DIR) systems and demonstrates strong generalization capabilities for the DIR task, offering a promising solution for handling variability in real-world data.

Abstract

The rapid evolution of intelligent document processing systems demands robust solutions that adapt to diverse domains without extensive retraining. Traditional methods often falter with variable document types, leading to poor performance. To overcome these limitations, this paper introduces a text-graphic layer separation approach that enhances domain adaptability in document image restoration (DIR) systems. We propose LayeredDoc, which utilizes two layers of information: the first targets coarse-grained graphic components, while the second refines machine-printed textual content. This hierarchical DIR framework dynamically adjusts to the characteristics of the input document, facilitating effective domain adaptation. We evaluated our approach both qualitatively and quantitatively using a new real-world dataset, LayeredDocDB, developed for this study. Initially trained on a synthetically generated dataset, our model demonstrates strong generalization capabilities for the DIR task, offering a promising solution for handling variability in real-world data. Our code is accessible on GitHub.

LayeredDoc: Domain Adaptive Document Restoration with a Layer Separation Approach

TL;DR

A text-graphic layer separation approach that enhances domain adaptability in document image restoration (DIR) systems and demonstrates strong generalization capabilities for the DIR task, offering a promising solution for handling variability in real-world data.

Abstract

The rapid evolution of intelligent document processing systems demands robust solutions that adapt to diverse domains without extensive retraining. Traditional methods often falter with variable document types, leading to poor performance. To overcome these limitations, this paper introduces a text-graphic layer separation approach that enhances domain adaptability in document image restoration (DIR) systems. We propose LayeredDoc, which utilizes two layers of information: the first targets coarse-grained graphic components, while the second refines machine-printed textual content. This hierarchical DIR framework dynamically adjusts to the characteristics of the input document, facilitating effective domain adaptation. We evaluated our approach both qualitatively and quantitatively using a new real-world dataset, LayeredDocDB, developed for this study. Initially trained on a synthetically generated dataset, our model demonstrates strong generalization capabilities for the DIR task, offering a promising solution for handling variability in real-world data. Our code is accessible on GitHub.
Paper Structure (20 sections, 2 equations, 5 figures, 1 table)

This paper contains 20 sections, 2 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Document Image Restoration
  4. Domain Adaptation for Document Analysis Systems
  5. Methodology
  6. Natural Image Restoration
  7. Layer Separation in Documents
  8. Learning Objectives
  9. Related Insights:
  10. Experiments
  11. Datasets
  12. Training Dataset
  13. LayeredDocDB
  14. Evaluation Metrics
  15. PSNR-Color:
  16. ...and 5 more sections

Figures (5)

  • Figure 1: General pipeline of LayeredDoc. We consider the standard architecture of image restoration models and propose to output two image layers instead of the standard single image in those methods. The first layer aims to output the text parts of the document, while the second layer aims to output the overlaid objects.
  • Figure 2: Visual illustration of the layer separation done by our 6-channel LayeredDoc model vs the ground truth
  • Figure 3: Some examples of our manually crafted LayeredDocDB dataset, illustrating layer 0, layer 1 and the merged noisy image with both layers.
  • Figure 4: Comparison between our LayeredDoc and the DocRes zhang2024docres approach. Our propose framework preserves the color in layer 1 as opposed to DocRes which in comparition puts the objects in gray scale.
  • Figure 5: Comparison between the proposed LayeredDoc model and the standard Restormer zamir2022restormer approach.