Complex Layout Classification in the Wild: A Low-Resource Approach with Layout-Preserving Augmentations

Technical Summary: Complex Layout Classification in the Wild

Problem Statement
Document layout analysis is a prerequisite for downstream tasks such as OCR, automated transcription, and reading-order reconstruction. While modern documents (e.g., PDFs, invoices) often follow clean "Manhattan" grid structures, historical documents—particularly medieval manuscripts and early printed books—exhibit highly irregular, complex layouts. These layouts feature non-rectangular text regions (e.g., L-shaped, C-shaped, O-shaped, and "Talmudic" multi-column arrangements) separated by distinct geometric boundaries.

Current state-of-the-art layout analysis relies heavily on large-scale annotated benchmarks (e.g., PubLayNet, DocLayNet) and deep learning models trained on standardized, English-language documents. These methods struggle in low-resource historical settings for three primary reasons:

1. Data Scarcity: Historical collections lack large-scale annotated datasets.
2. Structural Complexity: Standard pixel-based segmentation often fails to distinguish semantically distinct text blocks in complex configurations (e.g., failing to separate a main text from a surrounding commentary in a C-shape).
3. Overfitting: Models trained on small datasets tend to overfit to incidental textual details (font, specific content) rather than learning the global geometric arrangements that define the layout class.

Methodology
The authors propose a low-resource approach centered on a novel training strategy that prioritizes separator geometry over textual content.

1. Dataset Curation (CLC):

   • The authors curated the CLC dataset, a specialized corpus of 155 high-resolution Hebrew pages from the National Library of Israel and various printed books.
   • The dataset is manually annotated into eight distinct layout classes based on separator regions: C (and its mirror), L (and its mirror), U (and its inverted mirror), O, and Y (a catch-all for complex multi-component layouts).
   • Annotations were verified by an expert paleographer to ensure ground-truth accuracy.

2. Layout-Preserving Augmentations:
   To overcome data scarcity without introducing label noise, the authors developed a task-specific augmentation pipeline designed to suppress textual details while preserving the structural "skeleton" of the page:

   • Narrow Anisotropic Gaussian Masking: A multi-oriented anisotropic Gaussian filter is applied to the document image. By using a narrow standard deviation ($\sigma_x=1$) and filtering at horizontal ($\theta=0$) and vertical ($\theta=\pi/2$) orientations, the method generates a response map that highlights separator regions (which are continuous lines) while suppressing high-frequency textual noise. The resulting mask inverts non-separator pixels, forcing the model to ignore font and content.
   • Reflection-Induced Label Transformations: Horizontal and vertical reflections are applied to enrich the training distribution. Crucially, the authors define a closed label set where reflections map asymmetric classes to their counterparts (e.g., $C \leftrightarrow \bar{C}$, $L \leftrightarrow \bar{L}$) or to the $Y$ class, ensuring label consistency. Rotations and shears are explicitly excluded as they alter the structural semantics of the layout.

3. Model Architecture:

   • The system employs a ConvNeXt-Tiny backbone, chosen for its inductive biases (locality, translation equivariance, hierarchical composition) which align well with separator-driven geometry.
   • The model is fine-tuned using a staged approach: first training the classifier head, then unfreezing the backbone with a lower learning rate to prevent overfitting.
   • Input images are resized to $224 \times 224$ and binarized using Otsu's thresholding.

4. Active Learning Protocol:
   The authors evaluated label efficiency using an active learning loop with three acquisition functions (Random, Entropy, and Margin) to determine how few labeled samples are required to reach high performance.

Key Contributions

1. CLC Dataset: A new, compact, high-resolution dataset of 155 Hebrew pages with eight fine-grained, separator-defined layout classes, addressing the lack of resources for non-rectangular historical layouts.
2. Novel Augmentation Strategy: The introduction of Narrow Anisotropic Gaussian Masking, which compels the model to learn global geometric arrangements by suppressing incidental textual details, combined with reflection-induced label transformations to expand the training distribution while maintaining semantic consistency.
3. Systematic Evaluation: A comprehensive evaluation under severe data scarcity, including backbone comparisons, ablation studies, external blind-corpus testing, and active learning analysis.

Results

• Backbone Comparison: On the 31-page test set, ConvNeXt-Tiny achieved the best performance with 90% accuracy and a macro-F1 of 0.88. This significantly outperformed EfficientNet-B0, VGG16, ResNet-18, and ViT-B/16 (which scored $\le$ 45% accuracy), demonstrating that CNNs with appropriate inductive biases are superior to Transformers in this low-resource, geometry-focused regime.
• Ablation Study:
  • Binary only: 32% accuracy (indicating simple binarization is insufficient).
  • Binary + Flips: 45% accuracy.
  • Binary + Masking: 68% accuracy (confirming the task is separator-driven).
  • Full Pipeline (Masking + Flips): 90% accuracy.
  • The study confirmed that masking acts as a structure-preserving regularizer, while flips provide necessary diversity.
• External Blind Evaluation: The model was tested on a subset of 36,215 book folders from the National Library of Israel (unseen data). On a manually verified high-confidence subset (160 samples), the model achieved 83.75% accuracy, demonstrating strong generalization to real-world, uncurated archives.
• Active Learning: The full augmentation pipeline significantly improved label efficiency. Under an entropy acquisition strategy, the full pipeline achieved a macro-F1 Area Under the Learning Curve (AULC) of 0.61, compared to 0.30 for baseline methods.

Significance and Claims
The paper claims that robust classification of complex historical layouts is achievable under severe annotation scarcity by shifting the learning focus from textual content to global geometric structure. The authors argue that standard data augmentation (e.g., rotation, blur) is insufficient for this domain; instead, domain-aware augmentations that explicitly preserve separator topology are critical.

The work demonstrates that:

1. Layout classification for historical documents can be decoupled from text recognition, serving as a reliable routing step for downstream pipelines.
2. The proposed "separator-preserving" strategy allows models to generalize to unseen, non-rectangular layouts without massive-scale annotation.
3. The CLC dataset and the proposed methodology provide a targeted baseline for future research in low-resource historical document analysis, specifically for non-Latin scripts and complex manuscript structures.

The authors remain modest regarding the current scope, noting that the approach is currently limited to printed Hebrew pages and that future work is needed to extend these methods to handwritten manuscripts and integrate script-type recognition.