A Survey on Change Detection Techniques in Document Images

Abhinandan Kumar Pun, Mohammed Javed, David DoermannarXiv 2023 · DOI · arXiv

Abstract: The problem of change detection in images finds application in different domains like diagnosis of diseases in the medical field, detecting growth patterns of cities through remote sensing, and finding changes in legal documents and contracts. However, this paper presents a survey on core techniques and rules to detect changes in different versions of a document image. Our discussions on change detection focus on two categories -- content-based and layout-based. The content-based techniques intelligently extract and analyze the image contents (text or non-text) to show the possible differences, whereas the layout-based techniques use structural information to predict document changes. We also summarize the existing datasets and evaluation metrics used in change detection experiments. The shortcomings and challenges the existing methods face are reported, along with some pointers for future research work.

The Challenge of Document Change Detection

While change detection is a mature field in remote sensing (tracking city growth) and medical imaging (diagnosing disease progression), it has received surprisingly little attention in the context of document images.

The challenge isn't just comparing pixels. Documents are semi-structured; a single added word can cause a "cascading shift," moving every subsequent line and paragraph. Furthermore, real-world documents are often degraded, skewed, or contain non-textual elements like handwritten signatures, stamps, and logos. This survey formalizes the current state of the art, identifying which techniques can reliably distinguish between a meaningful content change and mere scanning noise or layout reformatting.

Comparison of two document versions (a) and (b) highlighting the difficulty of identifying content changes amidst layout shifts.

Methodologies

The survey divides the field into two primary methodologies: content-based and layout-based.

Content-Based Change Detection

  • OCR-Based: These methods convert images to text and use string matching algorithms like Longest Common Subsequence (LCS) or UNIX-style diff. While intuitive, they fail if the OCR engine misreads characters or if the document contains handwriting.
  • Feature-Based (SIFT): Using Scale-Invariant Feature Transform (SIFT) allows researchers to match local descriptors between images. This is more robust to "noise" like blur or skew than raw pixel comparison.
  • CIPDEC (Content Integrity of Printed Documents): A specialized approach using Error Correcting Codes (ECC) to detect pixel-level tampering in printed and then scanned documents.
  • VisualDiff++: A sophisticated pipeline that combines skew correction, page alignment, and dense SIFT matching to find the "longest feasible path" of matching components.
A typical pipeline for document verification and modification checking, illustrating the generation and verification stages.

Layout-Based Change Detection

These methods ignore the specific text and focus on the "skeleton" of the document. They use distance measures (edit distance, cluster distance, bitmap distance) to rank how similar two page structures are. The paper highlights LayoutLM, a recent framework that models the interaction between text and layout information using deep learning.

Benchmarks and Datasets

The paper identifies two primary benchmarks used in the field:

  • Enron Revision Dataset: Derived from the Enron email corpus, this contains roughly 150 one-page documents with 4–41 additions or deletions. It is often used for controlled experiments with varying DPI and blur levels.
  • Tobacco Revision Dataset: A more challenging, real-world dataset featuring nearly identical documents from tobacco industry archives. These include complex layouts, handwritten notes, signatures, and logos.

Key Insights: The OCR Bottleneck and the Domino Effect

Traditional OCR-based change detection is surprisingly fragile. If a scan is low-quality or uses a complex script, OCR error rates can exceed 70%, leading to massive "false positive" changes.

A major insight is the difficulty of "cascading shifts." Adding a single sentence in a two-column layout might push text from the bottom of the left column to the top of the right. Simple pixel or line-by-line comparisons fail here; the algorithm must understand the reading order. Consequently, image-based features (SIFT) often outperform OCR because they don't rely on character recognition and can detect changes in pen strokes or logos that OCR ignores.

Performance and Future Directions

The authors summarize performance metrics from seminal papers, noting that VisualDiff++ stands out with 91.1% accuracy on the Tobacco dataset, significantly outperforming basic OCR-diff methods (which hovered around 49.8%). SIFT-based approaches also achieved a superior Equal Error Rate (EER) of 0.1 compared to 0.147 for OCR-based methods.

Future research must address language barriers (moving beyond Latin scripts), semantic understanding (distinguishing formatting changes from fraudulent substitutions), and the detection of changes in non-textual elements like signatures and stamps.

Summary updated

Abstract

The problem of change detection in images finds application in different domains like diagnosis of diseases in the medical field, detecting growth patterns of cities through remote sensing, and finding changes in legal documents and contracts. However, this paper presents a survey on core techniques and rules to detect changes in different versions of a document image. Our discussions on change detection focus on two categories -- content-based and layout-based. The content-based techniques intelligently extract and analyze the image contents (text or non-text) to show the possible differences, whereas the layout-based techniques use structural information to predict document changes. We also summarize the existing datasets and evaluation metrics used in change detection experiments. The shortcomings and challenges the existing methods face are reported, along with some pointers for future research work.
BibTeX
@misc{pun2023survey,
  author = {Abhinandan Kumar Pun and Mohammed Javed and David Doermann},
  title = {{A Survey on Change Detection Techniques in Document Images}},
  year = {2023},
  eprint = {2307.07691},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.07691},
}
Figure