DocRevive: A Unified Pipeline for Document Text Restoration

Kunal Purkayastha; Ayan Banerjee; Josep Llados; Umapada Pal

cs.CVApr 11, 2026

DocRevive: A Unified Pipeline for Document Text Restoration

Authors: Kunal Purkayastha, Ayan Banerjee, Josep Llados, Umapada Pal

Organizations: Computer Vision Center · Indian Statistical Institute, Kolkata

Abstract

In Document Understanding, the challenge of reconstructing damaged, occluded, or incomplete text remains a critical yet unexplored problem. Subsequent document understanding tasks can benefit from a document reconstruction process. In response, this paper presents a novel unified pipeline combining state-of-the-art Optical Character Recognition (OCR), advanced image analysis, masked language modeling, and diffusion-based models to restore and reconstruct text while preserving visual integrity. We create a synthetic dataset of 30{,}078 degraded document images that simulates diverse document degradation scenarios, setting a benchmark for restoration tasks. Our pipeline detects and recognizes text, identifies degradation with an occlusion detector, and uses an inpainting model for semantically coherent reconstruction. A diffusion-based module seamlessly reintegrates text, matching font, size, and alignment. To evaluate restoration quality, we propose a Unified Context Similarity Metric (UCSM), incorporating edit, semantic, and length similarities with a contextual predictability measure that penalizes deviations when the correct text is contextually obvious. Our work advances document restoration, benefiting archival research and digital preservation while setting a new standard for text reconstruction. The OPRB dataset and code are available at Hugging Face and Github respectively.

Figures & tables

Figure 1: Feature completeness across document datasets: OPRB is the only benchmark that jointly provides degradation-aware labels, paired clean/degraded pages, word-level supervision, and restoration-oriented structure, making it more suitable for document restoration than standard layout-only benchmarks.

Figure 2: Sample visualization of occluding patches over a document page. Magnified views highlight representative degradations such as stamps, scribbles, black ink, and other occlusions that partially or fully obscure the text.

Figure 3: t-SNE comparison across document datasets: OPRB occupies both distinct and shared regions in the shared feature space, indicating that it captures degradation patterns and restoration-relevant document conditions not represented by standard clean-layout benchmarks.

Figure 4: PCA variance across datasets: OPRB exhibits broad variation in its document appearance and degradation patterns, which is important for evaluating restoration methods under diverse conditions.

Figure 5: DocRevive: a unified model architecture for the proposed document restoration pipeline. Given a degraded document image, the framework first performs OCR, followed by occlusion-aware blank region extraction to localize missing text areas. The surrounding word context is then grouped and converted into formatted prompt tokens, which are processed by the RoBERTa masked language model to predict the missing content. In parallel, style patches from nearby regions guide the text-editing branch, while text inpainting helps recover occluded neighboring text regions. Finally, postprocessing integrates the edited patches, cleans the gaps between lines, and generates a restored document image.

Figure 6: Visualization of the proposed framework on a real scanned document degraded using whiteboard marker ink. The document image in Left: was destroyed using concentrated ink, whereas the document in Right: was destroyed using diluted ink.

Method	Opaque Classes			Transparent Classes		Special	Average
	B.Ink	Burnt	Whitener	Dust	Stamp	Scribble	Avg.
YOLO26n [ 30 ]	93.701	92.254	90.880	71.175	86.802	86.103	86.819
YOLOv8n [ 46 ]	99.452	99.498	99.483	90.498	99.500	99.270	97.950
YOLOv8s [ 46 ]	99.493	99.500	99.496	91.895	99.487	99.442	98.219
YOLOv9c [ 41 ]	99.500	99.500	99.497	90.675	99.500	99.431	98.017
YOLOv10b [ 40 ]	99.490	99.487	99.454	92.153	99.079	99.288	98.159

Table 1: Class-wise occlusion detector ablation after removing the Mixed class. Values report mAP50 (%). Bold texts represent best results whereas underlined texts represent second best results.

Model	UCSM $\uparrow$	ED $\downarrow$	CER $\downarrow$	Exact Match (%) $\uparrow$
RoBERTa-large (fill-mask) [ 22 ]	0.6102	8.2662	1.3762	21.60
Qwen3-0.6B (causal LM) [ 38 ]	0.1240	7.0928	1.0807	0.15

Table 2: RoBERTa [ 22 ] vs. Qwen3 [ 38 ] text prediction ablation on a 6000-sample validation subset. Bold represent best results whereas underlined represent second best results. Higher UCSM and exact-match are better, while lower ED and CER are better.

Sample Set	UCSM $\uparrow$		Exact Match (%) $\uparrow$		ED $\downarrow$
	RoBERTa	Qwen3	RoBERTa	Qwen3	RoBERTa	Qwen3
0-1000	0.6131	0.1221	22.30	0.10	8.4530	7.3240
1001-2000	0.6201	0.1243	22.40	0.10	8.2380	7.0400
2001-3000	0.6093	0.1154	21.50	0.40	8.2000	7.0470
3001-4000	0.6175	0.1335	23.00	0.20	8.2210	7.0370
4001-5000	0.6046	0.1317	19.90	0.10	8.3670	7.0380

Table 3: Bucket-wise comparison between fine-tuned RoBERTa [ 22 ] and Qwen3 [ 38 ] on a validation fold. Each bucket contains 1000 consecutive expanded validation samples. Bold represent best results whereas underlined represent second best results.

Method and Variant	Precision	Recall	F1-Score	mIoU
LinkNet-ResNet18 [ 8 ]	89.46	92.52	90.97	74.00
LinkNet-ResNet34 [ 8 ]	87.38	91.07	89.18	74.00
LinkNet-ResNet50 [ 8 ]	88.58	93.56	91.00	74.00
DB-ResNet50 [ 21 ]	89.33	95.27	92.20	76.00
DB-MobileNetV3L [ 21 ]	88.60	93.80	91.13	74.00
FAST-tiny [ 11 ]	90.68	94.63	92.61	78.00

Table 4: Text detection performance on OPRB dataset. Bold texts represent best results whereas underlined texts represent second best results. Here DB means Differentiable Binarization

Method	Accuracy	1-NED	Confidence
CLIPOCR [ 42 ]	94.55	97.71	96.34
DATR [ 29 ]	89.97	96.19	90.94
PARSEQ [ 5 ] (Our Baseline)	95.42	98.03	95.85

Table 5: Text recognition performance on OPRB dataset. Bold represent best whereas underlined represent second best results.

Type	SSIM $\uparrow$	FID $\downarrow$	UCSM $\uparrow$
Black Ink	0.6844	19.10	0.6667
Burnt	0.6859	20.50	0.9631
Dust	0.7230	20.08	0.5138
Scribble	0.6748	19.41	0.6352
Stamp	0.6329	38.11	0.3717
Whitener	0.7560	14.36	0.5104

Table 6: Performance comparison across different occlusion types on DocRevive. Higher SSIM and UCSM indicate better perceptual quality, while lower FID indicates better distribution similarity.

\displaystyle\mathrm{Token}_{\alpha}=

\displaystyle\texttt{[K=}M_{\alpha}\texttt{] }\mathrm{PreText}_{\alpha}\;\texttt{<mask>}\;\mathrm{PostText}_{\alpha}

Algorithm 2 Prompt Token Generation for Emitted Blanks

Figure 7: Visualization of the proposed framework on a few OPRB document images for different types of occlusions. [Zoom in for better visualization]

Figure 8: Visualization of the state of the art industry models on a few OPRB document images for different types of occlusions. [Zoom in for better visualization]

Example 1: Semantic Blindness Context: “We evaluate the on three benchmark datasets.” GT: “proposed method”
	Prediction	ED	$S_{\text{edit}}$	$S_{\text{sem}}$	$S_{\text{len}}$	$E_{\text{ctx}}$	UCSM
	proposed method (exact match)	0	1.000	1.000	1.000	0.000	1.000
	proposed methoc (OCR typo)	1	0.933	0.665	1.000	0.000	0.853
	suggested approach (synonym)	13	0.278	0.859	0.833	0.000	0.584
	random variables (hallucination)	15	0.062	0.523	0.938	0.000	0.313
Example 2: Length Blindness Both predictions have ED = 5, but severity is opposite

Table 7: Three blind spots of Edit Distance that UCSM addresses.

S_{\text{edit}}

S_{\text{sem}}

, and

S_{\text{len}}

are defined in Eqs. ( 5 )–( 7 ),

E_{\text{context}}

in Eq. ( 9 ), UCSM in Eq. ( 10 ). Semantic embeddings are computed with all-MiniLM-L6-v2, context predictability with GPT-2.

Figure 9: Graphical visualization of effectiveness of UCSM over Edit Distance [Zoom in for better visualization]

Method	SSIM $\uparrow$	PSNR $\uparrow$	FID $\downarrow$
DocDiff	0.6620	16.46	43.25
GSDM	0.6634	16.58	53.78
NAFNet	0.6663	17.42	67.57
DocRevive (Ours)	0.6814	17.23	9.00

Table 8: Overall comparison on the OPRB subset. Best results in bold . Second best in underline .

	SSIM $\uparrow$						PSNR $\uparrow$
Method	Blk. Ink	Scrib.	Dust	Stamp	Burnt	Whit.	Blk. Ink	Scrib.	Dust	Stamp	Burnt	Whit.
DocDiff	0.6792	0.6726	0.6416	0.6538	0.6748	0.6500	15.90	16.90	16.45	16.12	16.91	16.51
GSDM	0.6814	0. 6751	0.6458	0.6556	0.6742	0.6486	16.93	17.37	15.61	15.75	17.11	16.72
NAFNet	0.6790	0.6699	0.6490	0.6609	0.6805	0.6587	16.73	17.77	17.44	17.10	17.88	17.59
DocRevive	0.6816	0.6745	0.6833	0.6768	0.6866	0.6855	17.24	17.09	17.26	16.83	17.48	17.50

Table 9: Per-occlusion-type SSIM

\uparrow

and PSNR

\uparrow

on the OPRB. Best in bold . Second best in underline

Panel of Figure 10 — Figure 10: Qualitative comparison of document restoration methods across occlusion types.

Explore similar work

Sep 27, 2026

ENet-GP: Unified Document Image Restoration

Reliable document digitization in uncontrolled capture settings is challenging because real images exhibit multiple interacting degradations rather than a single isolated distortion. Documents thus captured are affected simultaneously by geometric distortions, like page warping, as well as photometric degradations such as non-uniform illumination, and blurring. However, most existing approaches address these factors independently and are evaluated on benchmarks containing only one distortion type, limiting their real-world applicability. We introduce GutenDoc, a large-scale dataset of high-resolution dense-text documents with physically grounded compound degradations. Using physics-based rendering, our dataset jointly models geometric warping and diverse photometric effects, enabling systematic evaluation under realistic capture conditions. We further propose a unified restoration framework that jointly corrects geometric and photometric distortions within a single-network and single-training setup, without the need for degradation-specific retraining or sequential inference passes. Extensive experiments show that our method remains competitive on established single-distortion benchmarks while substantially improving robustness under compound degradations, providing a practical solution for real-world document digitization.

Sujal Burad, Aakanksha, A. N. Rajagopalan +1

Indian Institute of Technology, Madras · Adobe, India

Image Corruption Robustness Image Restoration

Ask PDF

Aug 29, 2026

DocIntent: Answerability-Guided Agentic Restoration for Real-World Document Visual Question Answering

Real-world degradations such as blur, shadow, distortion, and moire patterns severely impair the document question-answering capabilities of Multimodal Large Language Models (MLLMs). Applying restoration tools before Visual Question Answering (VQA) is an intuitive solution. However, existing restoration approaches remain limited, as manually designing and executing restoration strategies is labor-intensive and requires domain expertise. Agentic restoration offers new possibilities for automation, yet existing frameworks primarily target natural images and pursue perceptual quality, overlooking that restoration should serve downstream tasks rather than optimize generic image quality metrics. To this end, we explore the value of agentic restoration for real-world degraded document VQA and propose DocIntent, a training-free Answerability-Guided Agentic Restoration framework. DocIntent first assesses question answerability, then identifies task-relevant degradations and selectively invokes restoration tools. A Comparison-Based Rollback mechanism validates each restoration step and reverts it when question-relevant evidence becomes less decipherable. The entire process requires no additional pretrained degradation classifier or image quality assessment model. Extensive experiments on the WildDoc benchmark show that DocIntent consistently improves the average score and consistency of different open- and closed-source MLLMs. The code and experimental data will be publicly available.

Zihan Huang, Shihang Wu, Junle Liu +4

South China University of Technology

Visual Question Answering Multimodal QA Image Restoration

Ask PDF

Oct 12, 2026

DDVR-Net: Degradation-Aware Decoupling and Visibility-Guided Restoration for Historical Document Recognition

Historical document analysis is crucial for cultural heritage preservation and low-resource language digitization. However, real historical documents often suffer from severe degradation, such as ink erosion and structural damage, causing mixed-visibility character sequences. Conventional OCR systems struggle in such evidence-limited scenarios, while contextual language models may over-correct visually clear characters. To address these challenges, we propose DDVR-Net, a degradation-aware decoupling and visibility-guided restoration framework for historical document recognition. DDVR-Net decouples intrinsic glyph structures from physical corruption at the feature level and predicts local visibility to estimate character reliability. A contextual decoder is selectively activated through a visibility-guided gate, restoring degraded characters while preserving clear visual recognition. Furthermore, we construct a low-resource multilingual benchmark covering Tibetan, Mongolian, and Uyghur scripts with 187000 document pairs, combining large-scale synthetic data and progressive training with real-world annotations to reduce the synthetic-to-real gap. Extensive experiments demonstrate that DDVR-Net consistently outperforms existing methods across three languages and diverse reading orders, providing an effective framework and benchmark for low-resource historical document recognition. Code and Benchmark: https://github.com/Shajiu/DDVR-Net

Jiu Sha, Yu Weng, Mengxiao Zhu

Institute of National Security, Minzu University of China, Beijing, China · School of Information Engineering, Minzu University of China, Beijing, China · School of Artificial Intelligence and Computer Science, North China University of Technology, Beijing, China

Optical Character Recognition Historical Document Processing Low-Resource Language Processing

Ask PDF