BookNet: Book Image Rectification via Cross-Page Attention Network

Shaokai Liu; Hao Feng; Bozhi Luan; Min Hou; Jiajun Deng; Wengang Zhou

cs.CVJan 29, 2026

BookNet: Book Image Rectification via Cross-Page Attention Network

Authors: Shaokai Liu, Hao Feng, Bozhi Luan, Min Hou, Jiajun Deng, Wengang Zhou

Organizations: Hefei University of Technology, Hefei 230009, China. · University of Science and Technology of China, Hefei, 230027, China.

Abstract

Book image rectification presents unique challenges due to its complex geometric distortions from binding constraints, where left and right pages exhibit distinctly asymmetric curvature patterns. However, existing single-page document image rectification methods fail to capture the coupled geometric relationships between adjacent pages in books. In this work, we introduce BookNet, the first end-to-end deep learning framework specifically designed for dual-page book image rectification. BookNet adopts a dual-branch architecture with cross-page attention mechanisms, enabling it to estimate warping flows for both \rev{individual pages and the book spread}, explicitly modeling how left and right pages influence each other. Moreover, to address the absence of specialized datasets, we present Book3D, a large-scale synthetic dataset for training, and Book100, a comprehensive real-world benchmark for evaluation. Extensive experiments demonstrate that BookNet outperforms existing state-of-the-art methods on book image rectification. \rev{Code and dataset are available at https://github.com/Liu-Shaokai/BookNet.

Figures & tables

Fig. 1: Rectification paradigm comparison. (a) Conventional single flow for individual pages. (b) Single flow fails on books. (c) Our multi-flow solution effectively rectifies books by predicting separate flows (left, right, full).

Fig. 2: Book3D synthetic dataset generation pipeline and representative samples. Left: Blender rendering workspace showcasing the 3D book modeling environment with randomized book proportions and camera poses. Right: Rendered book samples from diverse arXiv academic papers, demonstrating realistic geometric deformations under varied illumination conditions and viewing angles. Top row shows the rendered synthetic book images, while bottom row displays the corresponding ground truth arXiv paper images.

Fig. 3: Representative samples from the Book100 benchmark dataset illustrating diverse capture conditions and content types. Top row: Distorted book images captured under various real-world conditions exhibiting different deformation patterns, lighting variations, and viewing angles. Bottom row: Corresponding high-quality reference scans obtained using professional overhead document cameras, providing ground truth for evaluation.

Attribute	Coverage
Language	English, Chinese, German, French, Japanese, and others
Subject	Mathematics, Physics, Management, Engineering, Computer Science, Medicine, Biology, Humanities, History, and others
Content Type	Text, Figures, Equations, Tables, Code, Mixed
Capture Time	Morning, Afternoon, Evening
Environment	Indoor, Outdoor
Lighting	Natural, Artificial, Mixed

TABLE I: Book100 Dataset Diversity Statistics

Fig. 4: Overview of the proposed BookNet architecture. Given a distorted book image $\mathbf{I}_{d}$ containing both left and right pages, our method extracts features through a CNN backbone and Transformer encoder. The dual-branch decoder employs a two-stage architecture with cross-page attention mechanisms to process learned queries, generating the left flow $\mathbf{M}_{l}$ , the right flow $\mathbf{M}_{r}$ , and the full flow $\mathbf{M}_{f}$ for the left page, the right page, and the book spread respectively . During training, all three flows are supervised, while inference uses the full flow $\mathbf{M}_{f}$ for final rectification.

Method	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$	CER $\downarrow$	ED $\downarrow$
Distorted	0.28	39.63	0.96	0.0530	172.63
DewarpNet [ 18 ]	0.41	19.98	0.62	0.0478	153.03
DocTr [ 19 ]	0.47	18.10	0.59	0.0516	165.77
DocGeoNet [ 21 ]	0.48	17.84	0.59	0.0507	166.93
PaperEdge [ 24 ]	0.46	14.95	0.56	0.0472	148.10
DocTr++ [ 28 ]	0.40	28.41	0.65	0.0559	185.43

TABLE II: Quantitative Comparison on Book100 Benchmark

Fig. 5: Visual comparison of flow supervision strategies. Left: left and right flows. Middle: full flow only. Right: joint supervision (ours) achieves better gutter alignment.

Fig. 6: Qualitative comparison of document rectification methods on various book pages. Each row shows results from a different method: Distorted (input), DewarpNet [ 18 ] , PaperEdge [ 24 ] , UVDoc [ 57 ] , DocRes [ 58 ] , and BookNet (ours). Our BookNet demonstrates consistent rectification quality across all document types while preserving text readability and image details.

Fig. 7: Zoomed-in qualitative comparison. Each example shows the distorted input, the results of DewarpNet [ 18 ] , UVDoc [ 57 ] and BookNet, and the reference scan. The rectified images of BookNet show less distortions than the other rectification methods.

Trans.	Dual-Br.	Feat. Fusion	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$	CER $\downarrow$	ED $\downarrow$
Encoder	Decoder	Module
	✓	✓	0.47	13.75	0.55	0.0556	185.13
✓		✓	0.48	12.75	0.52	0.0545	171.90
✓	✓		0.45	14.96	0.62	0.0650	206.80
✓	✓	✓	0.48	12.42	0.53	0.0450	141.63

TABLE III: Ablation Study on Architectural Components

Cross-Page Attention	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$	CER $\downarrow$	ED $\downarrow$
w/o Cross-Page Attention	0.46	13.11	0.58	0.0604	194.30
w/ Cross-Page Attention	0.48	12.42	0.53	0.0450	141.63

TABLE IV: Ablation Study on Cross-Page Attention

Left	Right	Full	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$	CER $\downarrow$	ED $\downarrow$
✓	✓		0.47	14.43	0.60	0.0611	202.57
		✓	0.45	15.01	0.61	0.0566	186.13
✓	✓	✓	0.48	12.42	0.53	0.0450	141.63

TABLE V: Ablation Study on Flow Supervision Strategies

$\boldsymbol{(\lambda_{l},\lambda_{r},\lambda_{f})}$	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$
$(2,2,1)$	0.47	12.75	0.58
$(1,1,2)$	0.47	12.91	0.57
$(1,1,1)$	0.48	12.42	0.53

TABLE VI: Ablation Study on Loss Weights

Fig. 8: Representative VQA case on the distorted input, the DocRes result, and the BookNet result, evaluated with Qwen2.5-VL-7B.

Fig. 9: Visualization of cross-page attention. The query point at the upper-left corner of the input image attends to the gutter region of the right page , demonstrating learned geometric coupling across pages.

Fig. 10: Robustness to off-center layouts. BookNet produces accurate rectification even when the book spine is significantly off-center.

Resolution	MSSIM $\uparrow$	LD $\downarrow$	AD $\downarrow$	FPS $\uparrow$
288 $\times$ 288	0.48	12.42	0.53	30.87
384 $\times$ 384	0.47	12.43	0.56	25.54
512 $\times$ 512	0.47	12.29	0.56	10.40

TABLE VII: Ablation Study on Input Resolution

Version	Distorted	DocRes	BookNet (Ours)
Accuracy	20/30 (66.7%)	25/30 (83.3%)	28/30 (93.3%)

TABLE VIII: Quantitative VQA Evaluation

Document Image Understanding

Document image understanding turns page images into recognized text, layout, structured data, and answers grounded in visual evidence.

See all 148 papers

Rethinking Supervision in Document Unwarping: A Self-consistent Flow-free Approach
Oct 11, 2026Shaokai Liu, Hao Feng, Wengang Zhou
MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230027, China.
Rectification Reimagined: A Unified Mamba Model for Image Correction and Rectangling with Prompts
Dec 21, 2025Linwei Qiu, Gongzhe Li, Xiaozhe Zhang +2All-in-One Image Restoration
Tianmushan Laboratory, Beihang University, Hangzhou, China · State Key Laboratory of High-Efficiency Reusable Aerospace Transportation Technology, Beijing, China · School of Astronautics, Beihang University, Beijing, China +2
ENet-GP: Unified Document Image Restoration
Sep 27, 2026Sujal Burad, Aakanksha, A. N. Rajagopalan +1Image Corruption Robustness Image Restoration
Indian Institute of Technology, Madras · Adobe, India

Oct 11, 2026J/K move · Enter open · S save

Rethinking Supervision in Document Unwarping: A Self-consistent Flow-free Approach

Shaokai Liu, Hao Feng, Wengang Zhou

MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230027, China.

Open paper AskPDF arXiv

In recent years, the proliferation of smartphones has led to an upsurge in the digitization of document files via these portable devices. However, images captured by smartphones often suffer from distortions, thereby negatively affecting digital preservation and downstream applications. To address this issue, we introduce DRNet, a novel deep network for document image rectification. Our approach is based on three key designs. Firstly, we exploit the intrinsic geometric consistency inherent in document images to guide the learning process of distortion rectification. Secondly, we design a coarse-to-fine rectification network to leverage the representations derived from the distorted document image, thereby enhancing the rectification result. Thirdly, we propose a unique perspective for supervising the learning of rectification networks, where undistorted document images are employed for supervision, which is free of warping mesh as ground truth in existing methods. Technically, both low-level pixel alignment and high-level semantic alignment jointly contribute to the learning of the mapping relationship between deformed document images and distortion-free ones. We evaluate our method on the challenging DocUNet Benchmark dataset, where it sets a series of state-of-the-art records, demonstrating the superiority of our approach compared to existing learning-based solutions. Additionally, we conduct a comprehensive series of ablation experiments to further validate the effectiveness and merits of our method.

Abstract

Document Image Understanding

Explore similar work

Rethinking Supervision in Document Unwarping: A Self-consistent Flow-free Approach

Rectification Reimagined: A Unified Mamba Model for Image Correction and Rectangling with Prompts

ENet-GP: Unified Document Image Restoration