Source-linked AI summary
Text Line Segmentation of Historical Documents: a Survey
Laurence Likforman-Sulem, Abderrazak Zahour, Bruno Taconet
TL;DR
Historical documents remain difficult to segment because degradation, irregular layouts, and interacting lines complicate extraction, despite growing needs for electronic access and analysis. This paper surveys historical-document text-line segmentation methods, organizing them into six categories and comparing their suitability to different line structures. The survey finds no universal method: performance and applicability depend on document characteristics, local adaptations, and careful parameter tuning.
Problem
Historical documents are difficult to exploit electronically because low quality, complex layouts, and interfering lines make text-line segmentation an open research problem.
Method
The paper surveys text-line segmentation methods for historical documents and organizes them into six major categories.
Results
Projection-based, smearing, grouping, Hough-based, repulsive-attractive network, and stochastic methods address different degradations and writing irregularities through distinct adaptations.
Takeaways & Limitations
No universal segmentation method fits all historical documents; methods can generalize to similar documents when parameters are tuned to their characteristics.
Takeaways & Limitations
Quantitative performance evidence is usually limited to reduced document sets, and no general comparative study of the methods had been carried out.
Abstract
from arXiv · showhide
There is a huge amount of historical documents in libraries and in various National Archives that have not been exploited electronically. Although automatic reading of complete pages remains, in most cases, a long-term objective, tasks such as word spotting, text/image alignment, authentication and extraction of specific fields are in use today. For all these tasks, a major step is document segmentation into text lines. Because of the low quality and the complexity of these documents (background noise, artifacts due to aging, interfering lines),automatic text line segmentation remains an open research field. The objective of this paper is to present a survey of existing methods, developed during the last decade, and dedicated to documents of historical interest.
1. Introduction
Historical documents are difficult to segment because their layouts, image quality, writing forms, and line interactions vary substantially. Text-line segmentation supports indexing, alignment, word spotting, authentication, and other document-analysis tasks.
- Challenges: Historical documents have looser layouts, aging-related degradation, disturbing artifacts, narrow spacing, overlapping components, and highly variable writing.These characteristics make their physical structure harder to extract than that of modern printed documents.
- Motivation: Digitized historical collections can be enhanced from basic metadata and transcriptions to searchable image-linked document representations.Text/image alignment records correspondences between transcription words and image regions.
- Applications: Text-line extraction is a prerequisite in many top-down word-spotting workflows and contributes to authentication based on writer characteristics.Authentication can use line orientations, fluctuations, and writing shapes independently of document content.
- Scope: The survey reviews text-line segmentation efforts for historical documents and situates preprocessing among the required processing stages.Preprocessing may remove non-textual or verso elements and improve binarization before extraction.
2. Characteristics and representation of text lines
Historical text lines are difficult to characterize because their geometry, spacing, orientation, insertions, artifacts, and connected components vary. The survey represents lines either through separating paths or through groups of aligned physical units.
- Definitions: Text-line segmentation assigns one label to spatially aligned units such as pixels, connected components, or characteristic points.The approaches either search for separating lines or paths, or search for aligned physical units.
- Complexity factors: Baseline fluctuation, varying orientations, marginal insertions, preprocessing errors, and fragmented or merged strokes complicate line assignment.Broken components may no longer remain linked to the writing’s median baseline.
- Complexity factors: Narrow line spacing makes segmentation harder because ascenders and descenders occupy the gaps between adjacent lines.The lower baseline of one line can approach the upper baseline of the next.
- Representations: Separating paths define text-line strips whose enclosed regions receive a common line label.Paths may be straight, piecewise straight, or curved.
- Representations: Cluster and string representations group units by shared labels or spatial order, while baselines follow fluctuations and require complementary handling of disconnected and touching units.Clusters can contain pixels, connected components, or blocks; strings are ordered aligned units.
3. Text line segmentation
Preprocessing is often necessary before extracting text lines from historical document images. It targets non-textual elements, bleed-through, background variation, and binarization problems while requiring adaptation to each document.
- Preprocessing: Preprocessing removes background noise and non-textual elements so text-line extraction can operate on cleaner document images.Targets include bindings, fingers, holes, stains, stamps, seals, ornamentation, and decorated initials.
- Binarization: Preprocessing must be accurately adapted to the characteristics of each document before text-line extraction.The survey presents preprocessing as a preparatory stage rather than a uniform operation.
- Preprocessing: Bleed-through writing can be reduced with filtering, wavelet techniques, or combinations of verso and recto images.These methods specifically address unwanted textual elements from the reverse side.
- Binarization: Global thresholding is generally unsuitable for historical documents with inhomogeneous backgrounds, whereas local thresholding can partially overcome this difficulty.Background normalization can make a global threshold easier to find, while faint writing may still cause over- or under-segmentation.
3.2 Projection–based methods
Projection-based methods derive text-line structure from image profiles and related geometric cues. They are simple and useful for separated or lightly overlapping lines, but short, narrow, and heavily overlapping lines can weaken profile evidence.
- Projection profiles: Vertical projection profiles sum pixel values across each row to reveal vertical gaps between text lines.The profile curve can then be smoothed and analyzed for maxima and minima.
- Projection profiles: Projection profiles are relatively insensitive to writing fragmentation, but short lines produce low peaks and narrow or heavily overlapping lines may lack significant peaks.Profiles can instead use black-white transitions or connected-component counts.
- Adaptations: Skew-aware projection applies a Hough transform to estimate page orientation before projecting along that angle and delimiting lines by profile minima.Profile maxima provide an estimate of the number of lines, while low maxima can be discarded relative to the highest peak.
- Adaptations: A distance-scored minimum can serve as an anchor for deriving remaining segmentation points when text lines are regularly spaced.This approach was applied to printed World War II records with regular line spacing.
- Adaptations: RXY cuts alternate projections along the X and Y axes to build a hierarchical tree and locate cuts within white spaces.Thresholds derive inter-line or inter-block distances for regular printed layouts or well-separated handwriting.
3.3 Smearing methods
Smearing methods connect text pixels horizontally to reveal text-line regions, with variants for binary, grayscale, and fragmented historical documents.
- Run-Length Smoothing fills horizontal white gaps below a threshold, after which connected-component bounding boxes enclose text lines.
- Gradient accumulation adapts smearing to grayscale images and requires adjustments for character alignment and line justification in old printed documents.
- A fuzzy run-length matrix skips limited foreground pixels, then thresholding produces text-line fragments without ascenders and descenders.Its parameters must be accurately and dynamically tuned.
3.4 Grouping methods
Grouping methods build text-line alignments bottom-up from image units, using neighborhoods, consistency criteria, and quality measures to manage complex handwritten layouts.
- Bottom-up grouping aggregates pixels, connected components, blocks, or salient points into alignments using local and global consistency criteria.
- Complex handwritten documents require explicit seed selection, neighborhood definition, and conflict resolution because nearest-neighbor joining can cross text lines.
- Quality measures select among competing alignments and may assess unit count, component size, spacing, or alignment straightness.
- Perceptual grouping extends directionally elongated connected components from selected anchors, resolving conflicts using alignment quality and higher-order neighboring components.
- Church Register methods construct body baselines from skeleton minima points, using angular neighborhoods initially and rectangular neighborhoods thereafter.
3.5 Methods based on the Hough transform
Hough-transform methods hypothesize line alignments in parameter space and validate them in the image domain, accommodating varying directions and handwritten fluctuations.
- Connected-component centroids are mapped to Hough cells (ρ, θ), where highly peaked cells indicate alignments containing many units.
- For fluctuating handwriting, alignments combine units from a primary cell and its surrounding cell structure rather than requiring perfect collinearity.
- The strongest alignment between two hypotheses is selected, then contextual internal and external neighbors validate or invalidate it to reject crossing alignments.
- Extracted line orientations correspond to primary-cell orientations in Hough space on an autograph of Miguel Angel Asturias.
- A moving image-domain window assigns remaining units after exhaustive Hough-domain searches begin from minima points in a left vertical strip.
3.6 Repulsive-Attractive network method
Repulsive-attractive networks iteratively adapt baseline positions directly on grayscale historical documents, using pixels and previously extracted baselines as opposing forces.
- The method constructs pseudo-baselines sequentially from top to bottom by attracting each baseline toward text pixels while repelling it from previously extracted baselines.
- Each baseline is initialized below the previously examined one, and the first begins in the blank space above the document.
- The resulting pseudo-baselines pass through word bodies, require similar line lengths, and were applied to ancient Ottoman archives and Latin texts.
3.7 Stochastic method
The stochastic method uses a probabilistic Viterbi algorithm and hidden Markov modeling to derive non-linear paths between text lines, including overlapping lines.
- A probabilistic Viterbi algorithm derives non-linear paths between overlapping text lines.The method was applied to modern Chinese handwriting, with the paper suggesting that the principle could extend to historical documents.
- Hidden Markov modeling extracts lines after the image is divided into small cells, each corresponding to an HMM state.
- The method produces segmentation paths that turn around obstacles rather than crossing overlapping components.
3.8 Processing of overlapping and touching components
Overlapping and touching components are detected and classified using geometric, alignment, contextual, or clustering cues, then separated with cuts, contour analysis, rules, or configuration dictionaries.
- Ambiguous components are identified when they overlap or touch across text lines, leaving no white space between lines.
- Detection can use component size, membership in multiple alignments, or absence from any alignment.
- Detected components are classified as upper- or lower-line overlaps, or touching components requiring decomposition into two or more parts.
- Vertical separation ranges from rough horizontal cuts to contour analysis guided by typical configurations.
- Other methods use alignment conflicts, density and proximity rules, clustering, profile-based anchor points, skeletons, contact contours, or connected-component centers of gravity.Center-of-gravity splitting works well for single characters but may fail for words or larger units.
3.9 Non Latin documents
Non-Latin historical documents require segmentation methods adapted to script-specific writing structures, spacing, diacritics, and connected components.
- Latin documents contain inter-line dots, ascenders, and descenders, while Arabic is connected and cursive with loops and diacritical points.
- Ancient Arabic pages may have faint inter-line spaces, consecutive lines connected by characters, overlapping lines, and fluctuating baselines.
- Zahour et al. recursively cluster and link blocks by y coordinates to generate line separators for ancient Arabic documents.
- Hebrew squared writing is suited to connected-component segmentation because most characters form single connected components.
- The Hough-based method produces text-line segmentation for Hebrew scrolls, as illustrated in Figure 17.
- Table 1 presents methods suitable for historical documents, including piecewise projection and k-means clustering for overlapping or touching lines.
4. Discussion and concluding remarks
The survey organizes historical text-line segmentation methods and compares their strengths, adaptations, and limitations across degradations, writing irregularities, and document types.
- The survey identifies six categories: projection-based, smearing, grouping, Hough-based, repulsive-attractive network, and stochastic methods.Most methods address some degradations and writing irregularities specific to historical documents.
- Projection, smearing, and Hough-based methods require local adaptations to handle proximity, overlaps, touching strokes, fluctuating lines, and fragmentation.
- The stochastic method is conceptually more robust, but careful implementation and initialization are required.
- Projection and Hough-based methods suit clearly separated lines, whereas stochastic paths avoid overlapping components and grouping methods can accumulate early errors.
- Baseline-based representations fit fluctuating lines, while stochastic methods can generate non-linear paths and shortest cutting paths for touching characters.
- Pixel-based methods tolerate fragmentation, but fragments far from baselines can generate spurious points that reduce accuracy or produce wrong representations.
- Quantitative comparisons are limited because evaluations often use reduced datasets, difficult ground truth, non-isolated segmentation tasks, and document-specific tuning.
- No universal method fits all historical documents; methods can generalize to similar documents through tuning for script size, stroke width, and spacing.