# Document Layout Analysis: A Comprehensive Survey > A survey of 79 document layout analysis studies organized around preprocessing, bottom-up/top-down/hybrid analysis strategies, and three tiers of evaluation metrics. - Authors: Bevaya AI Team - Published: 2026-03-14 - Paper: Published (ACM 2019) - Source: Bevaya Labs, --- ## TL;DR BinMakhashen and Mahmoud survey 79 document layout analysis studies, organizing the field around a general DLA framework consisting of preprocessing, analysis parameter estimation, layout analysis (bottom-up, top-down, hybrid), post-processing, and performance evaluation. Bottom-up methods dominate the literature (61% of reviewed papers) due to their flexibility with complex layouts, while top-down methods remain useful for structured documents. The survey predates the shift to deep learning in DLA (PubLayNet, LayoutLM, DETR-based detectors all appeared in 2019-2020), making it a useful historical reference for the classical foundations the field built upon. --- ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Systematic}}$ This is a survey paper. The headline contribution is a unifying DLA framework that organizes the field into preprocessing, analysis strategies (bottom-up, top-down, hybrid), post-processing, and evaluation. The paper reviews 79 studies and categorizes them by strategy, document type, language, and layout complexity. Most of the paper is devoted to organizing and explaining prior work rather than introducing new methods or datasets. **Secondary:** $\Psi\_{\text{Evaluation}}$ The survey dedicates a full section to performance evaluation, distinguishing three levels of metrics: pixel-level (PLEF), region-level (REF, IoU), and customizable (CEF, used in ICDAR competitions). It provides formal definitions for each and discusses their strengths and limitations. ## What is the motivation? The authors identify several gaps that motivate the survey: 1. **No unified framework.** DLA algorithms vary widely in their processing pipelines, making it difficult to compare methods or understand where a new contribution fits. Earlier surveys focused on narrow topics (texture-based methods, skew detection, printed-document DLA) rather than the full pipeline. 2. **Historical documents neglected.** Most prior surveys concentrated on contemporary printed documents, despite historical manuscripts presenting harder layout challenges (arbitrary layouts, degraded text, multi-script content). 3. **Evaluation inconsistency.** Many early DLA methods used subjective evaluation or incompatible metrics, hindering progress. The paper aims to consolidate the evaluation landscape. ## What is the novelty? The paper's primary contribution is the **general DLA framework** (Figure 2 in the paper), which decomposes DLA into five phases: 1. **Preprocessing:** Skew detection/correction and binarization. 2. **Analysis parameter estimation:** Static (predefined thresholds for structured layouts) vs. dynamic (data-driven measurements for heterogeneous documents). 3. **Layout analysis:** The core segmentation step, divided into three strategies: - **Bottom-up** (61% of 79 papers): Starts from pixels or connected components, grows to regions. Includes connected component analysis (Docstrum), texture analysis (Gabor, autocorrelation), machine learning (both non-deep and deep), Voronoi diagrams, and Delaunay triangulation. - **Top-down** (29%): Starts from page-level and splits into regions. Includes texture-based, RLSA, projection profile (X-Y cut), and whitespace analysis. - **Hybrid** (6%): Combines both strategies. 4. **Post-processing:** Optional refinement (clustering, morphological cleaning, human interaction). 5. **Performance evaluation:** Three tiers of metrics. The survey also adopts and extends Kise's **document layout taxonomy**, distinguishing six layout types: regular, Manhattan, non-Manhattan, multi-column Manhattan, arbitrary/complex, and overlapping. The extension to historical manuscripts with arbitrary layouts is the authors' contribution; the base categorization for printed documents originates from Kise [82]. ### Evaluation Framework Taxonomy The paper organizes DLA evaluation into three levels: **Pixel-Level Evaluation (PLEF):** Counts pixel matches between segmentation and ground truth. The match score between ground-truth region $G\_j$ and segmented region $R\_i$ over image foreground $I$ is: $$ \text{MS}(i, j) = \alpha \frac{\text{T}(G\_j \cap R\_i \cap I)}{\text{T}((G\_j \cup R\_i) \cap I)}, \quad \alpha = \begin{cases} 1, & \text{if } g\_j = r\_i \\\\ 0, & \text{otherwise} \end{cases} $$ where $\text{T}(\cdot)$ counts foreground pixels and $g\_j$, $r\_i$ are the labels of the ground-truth and segmented regions, respectively. The match table is aggregated into detection rate (DR) and recognition rate (RR) via weighted one-to-one, one-to-many, and many-to-one correspondences, then combined as $F\_{\text{measure}} = 2 \cdot \text{DR} \cdot \text{RR} / (\text{DR} + \text{RR})$. **Region-Level Evaluation (REF):** Uses IoU borrowed from computer vision. For each class $i$: $$ \text{IoU}\_i = \frac{G\_i \cap R\_i}{t\_i + (G\_i \cup R\_i) - (G\_i \cap R\_i)} $$ The mean IoU averages over $N$ classes ($\text{mIoU} = \frac{1}{N} \sum\_i \text{IoU}\_i$). A frequency-weighted variant scales each class by its pixel count $t\_i$. **Customizable Evaluation Framework (CEF):** Introduced by Antonacopoulos and Bridson (ICDAR 2007), this framework transforms regions into interval representations, establishes correspondence, then classifies errors as merge, split, miss, partial miss, or false detection. Each error type $i$ has a weighted error rate $ER\_i$ (the affected area multiplied by an application-specific weight $\alpha\_i$). The per-type weight is: $$ w\_i = \frac{(N - 1) \cdot ER\_i + 1}{N} $$ where $N$ is the number of error types considered. The overall success rate is then: $$ SR = \frac{\sum\_{i=1}^{N} w\_i}{\sum\_{i=1}^{N} \frac{w\_i}{1 - ER\_i}} $$ This design allows practitioners to customize error severity by adjusting $\alpha\_i$ for different analysis objectives (e.g., penalizing column merges more heavily for OCR applications). ## What experiments were performed? This is a survey; no new experiments are conducted. The paper collects and compares reported results from the literature across multiple axes: - **Strategy distribution:** Bottom-up (61%), top-down (29%), hybrid (6%), other (4%). - **Document type:** Printed (58%), handwritten (33%), other (9%). - **Language:** English (41%), Arabic (16%), German (9%), Latin (9%), French (8%), and others. - **Layout complexity:** Manhattan (37%), multi-column (35%), complex (28%). Table 1 compares skew detection methods by angle range, error, document type, and language. Table 2 catalogues 79 DLA algorithms by strategy, method type, document properties, and output. Table 4 collects quantitative results organized by evaluation metric. ## What are the outcomes/conclusions? The survey reaches several conclusions: 1. **Bottom-up dominates but is slow.** Connected-component-based analysis is the most flexible approach for complex layouts, but it carries higher computational cost than top-down methods. A complex layout can be analyzed top-down in 0.7 seconds with 78.5% success rate; bottom-up achieves higher accuracy at the cost of longer processing time. 2. **Deep learning is emerging.** The authors report that deep learning methods (FCNNs, U-Net variants) produced the strongest results among the surveyed papers, but require post-processing (clustering, morphological operations) and large training data. Transfer learning from ImageNet helps with convergence. 3. **Hybrid methods are underexplored.** Only 6% of reviewed papers use hybrid strategies, despite the complementary strengths of bottom-up (precision on complex layouts) and top-down (speed on structured layouts). 4. **Binarization is declining for modern methods** but remains critical for historical documents. Deep learning methods use full pixel intensities rather than binarized inputs. 5. **Evaluation needs standardization.** Many early methods used subjective evaluation. The CEF framework from ICDAR competitions offers the most comprehensive approach, but adoption is inconsistent. ### Limitations - **Temporal cutoff.** The survey was submitted in July 2018 and published in October 2019, so it predates the rapid adoption of deep learning in DLA. PubLayNet (2019), LayoutLM (2019), Mask R-CNN on documents, and DETR-based detectors are all absent. The "deep learning" methods discussed are early FCNNs and U-Nets, not the object detection or multimodal pre-training approaches that now dominate. - **Narrow language scope.** Despite noting the importance of language diversity, the survey focuses primarily on Latin and Arabic scripts. CJK, Indic, and other scripts receive minimal attention. - **No deep learning taxonomy.** The bottom-up/top-down/hybrid taxonomy does not cleanly accommodate modern object detection pipelines (Faster R-CNN, YOLO, DETR), which operate at neither the pixel/CC level nor the page-splitting level. - **Dataset coverage.** The dataset section (Table 3) lists only 12 datasets, all with fewer than 7K pages. The large-scale weakly-supervised datasets (PubLayNet at 360K, DocBank at 500K) that became widely adopted appeared shortly after this survey. --- ## Reproducibility ### Models Not applicable (survey paper). ### Algorithms Not applicable. The paper reviews but does not implement algorithms. ### Data The survey catalogues 12 DLA datasets (Table 3), including: - **Historical:** OHG (596 pages), DIVA-HisDB (150 pages, ICDAR 2017), Parzival (47 pages), George Washington GW20 (20 pages), Saint Gall (60 pages), IMPACT (7K pages) - **Contemporary:** UW-3 (1,600 pages), PRImA (305 pages), BCE-Arabic-v1 (1,833 pages), CENIP-UCCP (400 Urdu pages), LAMP (203 mixed pages), MAURDOR (2.5K pages) These are all small by modern standards. The survey does not discuss dataset licensing. ### Evaluation The paper reviews three evaluation frameworks (PLEF, REF, CEF) and collects quantitative results in Table 4, but does not propose new metrics or benchmarks. ### Hardware Not discussed. This is a survey, so no training or inference hardware is reported. ## BibTeX ```bibtex @article{binmakhashen2019document, title={Document Layout Analysis: A Comprehensive Survey}, author={BinMakhashen, Galal M. and Mahmoud, Sabri A.}, journal={ACM Computing Surveys}, volume={52}, number={6}, pages={1--36}, year={2019}, publisher={ACM}, doi={10.1145/3355610} } ```