D2Dewarp: Dual Dimensions Geometric Representation Learning Based Document Image Dewarping¶

Conference: CVPR 2026
arXiv: 2507.08492
Code: Available
Area: Self-Supervised Learning / Document Image Understanding
Keywords: Document Dewarping, Dual Dimension, Geometric Lines, UNet, HV Fusion

TL;DR¶

Ours introduces D2Dewarp, the first dewarping method to learn document geometric representations from both horizontal and vertical dimensions. A UNet with dual decoders predicts horizontal lines (top/bottom boundaries of text lines/tables) and vertical lines (left/right boundaries) separately. The HV Fusion Module integrates features from both directions via mixed attention. Additionally, the DocDewarpHV dataset, containing 114K images with dual-dimension annotations, is constructed to support this framework.

Background & Motivation¶

1. Background¶

Document Image Dewarping aims to restore curved or crumpled document images to a flat state, serving as a critical preprocessing step for OCR and document analysis. Existing methods are categorized into: (i) 3D coordinate regression (e.g., DewarpNet); (ii) 2D optical flow/displacement field methods (e.g., DocTr); (iii) geometric line-based methods (e.g., RDGR), which provide geometric constraints by predicting text line boundaries.

2. Limitations of Prior Work¶

Focus on horizontal lines only: Methods like RDGR utilize only horizontal geometric lines (upper/lower text boundaries), completely ignoring vertical structural information.
Neglected vertical distortion: Vertical bending is significant in scenarios like book gutters, table column boundaries, and multi-column layouts, yet no method explicitly models it.
Insufficient feature fusion: Even if both horizontal and vertical features are extracted, the lack of an effective cross-fusion mechanism prevents the full utilization of complementary information.
Lack of annotated data: Existing datasets (Doc3D, DocUNet) do not provide vertical line annotations, limiting the feasibility of dual-dimension learning.

3. Key Challenge¶

Document distortion is inherently a 2D spatial deformation, but existing methods rely on 1D (horizontal) geometric constraints, leading to fundamental information incompleteness.

4. Goal¶

To leverage geometric structural information from both horizontal and vertical dimensions simultaneously to guide document dewarping, supported by a specialized dataset.

5. Key Insight¶

Starting from dual-dimension geometric representation learning: learn structural features of horizontal and vertical lines separately, then fuse complementary information from both directions via an attention mechanism to generate high-quality deformation mappings.

Method¶

Overall Architecture¶

Document dewarping addresses how to "flatten" a distorted photograph. D2Dewarp's core premise is that document curvature is a 2D problem; horizontal text lines are insufficient when vertical table columns and binding lines are also warped. It utilizes a UNet architecture: an image enters a shared encoder to extract multi-scale features, then branches into two decoders—one for the H-Line Map (horizontal) and one for the V-Line Map (vertical). These decoders exchange information through the HV Fusion Module at intermediate layers. The model finally outputs a 2D backward mapping (displacement field), directing each distorted pixel to its correct position in a flat document. Geometric lines serve as both signals and constraints, anchoring abstract "deformation" to concrete structural boundaries. This training is enabled by the newly created DocDewarpHV dataset.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Distorted Document Image"] --> B["Shared Encoder<br/>Multi-scale Feature Extraction"]
    subgraph DEC["Dual Decoders (H/V Branches)"]
        direction TB
        C["H-Decoder<br/>Predict H-Line Map (Top/Bottom)"]
        D["V-Decoder<br/>Predict V-Line Map (Left/Right)"]
    end
    B --> DEC
    subgraph FUSE["HV Fusion Module"]
        direction TB
        E["Direction-Aware Pooling<br/>H along X-axis / V along Y-axis AvgPool"] --> F["Mixed Attention<br/>H↔V Cross Attention"]
        F --> G["Directional Self-Attention<br/>Restore Spatial Resolution"] --> H["Sigmoid Gated Reweighting<br/>Residual into Decoder Features"]
    end
    DEC --> FUSE
    FUSE --> I["2D Backward Mapping (Displacement Field)"]
    I --> J["Flattened Document Output"]
    K["DocDewarpHV Dataset<br/>H/V-Line + 3D + UV Labels"] -.Supervision.-> DEC

Key Designs¶

1. Dual Decoders: Decoupled Learning of H and V Geometric Lines

Previous methods like RDGR only predicted horizontal lines—the boundaries for text lines, table rows, or paragraphs—reflecting vertical bending. However, vertical structures like table columns and book gutters (V-Lines) also carry crucial curvature information. D2Dewarp splits the path after the shared encoder into H-Decoder and V-Decoder. This prevents geometric features from interfering within a single path, allowing each branch to focus on its specific directional bending pattern.

2. HV Fusion Module: Explicit Cross-Directional Feature Interaction

Since horizontal and vertical warping are physically coupled, an explicit exchange of context is required. Unlike simple concatenation, HV Fusion uses a four-step process: Direction-Aware Pooling (AvgPool along the X-axis for H and Y-axis for V to reduce redundant dimensions); Mixed Attention (cross-attention between pooled H/V features); X/Y Self-Attention to restore spatial resolution; and Sigmoid Gating to reweight and merge fused features back into the decoder. This "pool-then-cross" approach is more computationally efficient and directionally focused than standard global attention.

3. DocDewarpHV Dataset: Filling the Gap in Vertical Annotations

Existing datasets like Doc3D only provide horizontal annotations. Ours constructs DocDewarpHV using Blender to render flat documents onto curved 3D meshes, automatically extracting geometric H and V boundary lines. Each image includes 3D coordinates, UV mapping, H-Line maps, and V-Line maps. With 114,000 images (512×512), it surpasses Doc3D in both scale and annotation diversity, covering text-only, tabular, and mixed layouts.

Loss & Training¶

The total loss is a weighted sum of reconstruction loss and line prediction loss:

\[\mathcal{L} = \alpha \cdot \mathcal{L}_{\text{rec}} + \mathcal{L}_{\text{line}}\]

\(\mathcal{L}_{\text{rec}}\) (Reconstruction Loss): L1 loss calculating pixel-level differences between predicted backward mapping and GT displacement fields.
\(\mathcal{L}_{\text{line}}\) (Line Prediction Loss): Weighted BCE loss for H-Line and V-Line maps. Positive samples (line pixels) are weighted to mitigate extreme class imbalance (lines usually occupy < 5% of pixels).
Weight \(\alpha\): Hyperparameter to balance reconstruction and structural constraints.

Training config: Adam optimizer, learning rate 1e-4, batch size 16, 300 epochs. Trained on DocDewarpHV and tested on DocUNet, DIR300, and WarpDoc benchmarks.

Key Experimental Results¶

Main Results¶

Table 1: Comparison on DocUNet Benchmark (130 real distorted documents)

Method	MS-SSIM↑	LD↓	CER↓
DewarpNet	0.4735	8.39	0.4210
DocTr	0.5105	7.76	0.3561
DocGeoNet	0.5040	7.71	0.3806
RDGR	0.5224	7.61	0.3343
RecDocNet	0.5198	7.42	0.3482
D2Dewarp (Ours)	0.5387	7.18	0.3127

Table 2: Comparison on DIR300 Benchmark (300 documents)

Method	MS-SSIM↑	LD↓
DewarpNet	0.4868	8.98
DocTr	0.5241	7.94
RDGR	0.5356	7.63
D2Dewarp (Ours)	0.5521	7.28

Ablation Study¶

Table 4: Core Component Ablation (DocUNet Benchmark)

Configuration	MS-SSIM↑	LD↓
Baseline (Single Decoder + H-Line only)	0.5224	7.61
+ V-Line Branch (Dual Decoder, no fusion)	0.5298	7.42
+ Simple Concat Fusion	0.5315	7.36
+ HV Fusion Module (Full)	0.5387	7.18

Key Findings¶

Dual dimensions significantly outperform single dimension: Adding the V-Line branch alone (even without fusion) improves MS-SSIM from 0.5224 to 0.5298, confirming the importance of vertical geometric information.
Fusion mechanism is critical: The HV Fusion Module contributes an additional 0.72% MS-SSIM gain over simple concatenation, proving direction-aware design is superior to naive attention.
Consistently SOTA across three benchmarks: Ours leads in all metrics (MS-SSIM, LD, CER) on DocUNet, DIR300, and WarpDoc.
Substantial OCR improvement: CER decreased by 6.5% compared to RDGR on DocUNet, indicating that higher dewarping quality directly benefits downstream recognition.

Highlights & Insights¶

Addressing a neglected blind spot: While document distortion is 2D, prior works relied on 1D constraints. D2Dewarp identifies and solves this fundamental limitation.
End-to-end dual-dimension learning: The architecture ensures efficiency while achieving directional feature complementarity.
Dataset contribution: DocDewarpHV (114K images) is a vital contribution to the community, removing the bottleneck of missing vertical line labels.
Direction-aware attention: Pooling along corresponding axes before cross-attention is more efficient and geometrically meaningful than standard global attention.

Limitations & Future Work¶

Domain gap between 3D rendered data and real-world distributions may still limit generalization.
H-Line and V-Line definitions rely on regular structures (text, tables); performance might degrade for handwritten or highly irregular layouts.
Dual decoders increase parameter count and FLOPs by approximately 40%, potentially limiting mobile deployment.
HV Fusion is currently fixed at specific layers; future work could explore multi-scale fusion.

RDGR [Li et al.]: A baseline using horizontal text lines; D2Dewarp extends this from 1D to 2D.
DewarpNet [Das et al.]: First end-to-end deep learning method using 3D coordinates; D2Dewarp demonstrates that geometric lines can be more effective than raw 3D coordinates.
Doc3D [Das et al.]: Standard dataset for dewarping; DocDewarpHV improves upon it with V-Line labels and a larger scale.
Insight: Directional decoupling can be extended to other tasks, such as layout analysis (simultaneous horizontal/vertical cutting) or table recognition (modeling row and column lines separately).

Rating¶

Novelty: ⭐⭐⭐⭐ First systematic introduction of dual-dimension modeling.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across three benchmarks and detailed ablations.
Writing Quality: ⭐⭐⭐⭐ Clear structure with intuitive visualizations.
Value: ⭐⭐⭐⭐ Significant contribution via the DocDewarpHV dataset.