Beyond Pixel Reconstruction: Retrieval-Guided Glyph-Aware Restoration for Low-Resource Manchu Historical Documents

Ting Huang; Dongdong Wang; Mingqiu Liang; Siyang Lu

cs.CVSep 29, 2026

Beyond Pixel Reconstruction: Retrieval-Guided Glyph-Aware Restoration for Low-Resource Manchu Historical Documents

Authors: Ting Huang, Dongdong Wang, Mingqiu Liang, Siyang Lu

Organizations: School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China · College of Design, Construction, and Planning, University of Florida, Gainesville, FL, USA

Abstract

Historical Manchu documents preserve invaluable linguistic and cultural heritage, yet their digitization is hindered by severe degradations and the scarcity of paired training data. Existing document restoration methods primarily optimize pixel-level reconstruction, which can produce visually plausible results while failing to preserve the structural identity of Manchu glyphs. To address this limitation, we propose a retrieval-guided glyph-aware restoration framework that goes beyond pixel reconstruction by explicitly incorporating glyph-level structural knowledge. Our method retrieves relevant glyph exemplars to provide structural guidance during restoration and integrates this information into the reconstruction process, improving the recovery of degraded character structures under low-resource conditions. Extensive experiments on Manchu historical documents demonstrate that the proposed approach improves both image restoration quality and glyph-level fidelity compared with existing restoration methods. These results highlight the importance of incorporating character-aware structural priors for reliable restoration of low-resource historical documents.

Figures & tables

Figure 1: Overview of the proposed CRGR framework. CRGR reformulates restoration from pixel synthesis to retrieval-guided glyph recovery by progressively reducing candidate ambiguity through coarse retrieval and VLM-based verification.

Stage	Input	Transformation	Output
Coarse Glyph Retrieval	Masked glyph $x_{q}$ , gallery $\mathcal{G}$	Visual feature extraction + retrieval	Top- $K$ glyph candidates $\mathcal{C}$
Candidate Verification	Masked glyph $x_{q}$ , candidates $\mathcal{C}$	VLM-based vision-language matching	Reference glyph $g^{*}$
Glyph Restoration	Reference glyph $g^{*}$ , target location	Glyph alignment and regional replacement	Restored documents

Table 1: Data transformation across different stages of the proposed CRGR framework.

Figure 2: Qualitative comparison of full-page document restoration.

	Masked Regions					Intact Regions
Method	GRA $\uparrow$	CCR $\downarrow$	GSS $\uparrow$	GSS m $\uparrow$	GSS h $\uparrow$	GRA $\uparrow$	CCR $\downarrow$	GSS $\uparrow$	GSS m $\uparrow$	GSS h $\uparrow$
LDM	0.9354	0.0456	0.9966	0.9969	0.9959	0.9439	0.0448	0.9974	0.9976	0.9970
LaMa	0.8992	0.0684	0.9974	0.9977	0.9967	0.9589	0.0383	0.9997	0.9998	0.9997
AOT-GAN	0.8764	0.1008	0.9963	0.9965	0.9957	0.9418	0.0428	0.9970	0.9973	0.9967
GSDM	0.5627	0.0932	0.9971	0.9973	0.9966	0.5910	0.0688	0.9981	0.9982	0.9980
PixelHacker	0.9487	0.0380	0.9967	0.9970	0.9959	0.9283	0.0452	0.9966	0.9971	0.9960

Table 2: Quantitative comparison of glyph-oriented restoration performance on masked and intact regions.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE $\downarrow$	F1 $\uparrow$
LDM	34.671	0.996	0.004	4.875	0.999
LaMa	33.396	0.997	0.004	5.707	0.999
AOT-GAN	22.514	0.874	0.327	19.144	0.987
GSDM	20.464	0.936	0.123	25.787	0.987
PixelHacker	18.954	0.788	0.383	28.839	0.975
CRGR (Ours)	31.982	0.994	0.002	7.135	0.999

Table 3: Quantitative comparison of restoration performance.

Variant	GRA $\uparrow$	CCR $\downarrow$	GSS $\uparrow$	GSS m $\uparrow$	GSS h $\uparrow$
ConvNeXt	0.6186	0.2542	0.9977	0.9979	0.9969
Swin V2	0.6356	0.2288	0.9979	0.9980	0.9973
ConvNeXt + Qwen-VL	0.5763	0.2458	0.9976	0.9975	0.9980
ConvNeXt + InternVL	0.5678	0.2542	0.9976	0.9975	0.9980
Swin V2 + Qwen-VL	0.6356	0.2373	0.9980	0.9981	0.9976
Swin V2 + InternVL	0.6610	0.2203	0.9979	0.9980	0.9972

Table 4: Ablation study of CRGR variants. Bold : best result; underlined : second-best result.

Figure 3: Qualitative comparison of different CRGR variants.

$K$	GRA $\uparrow$	CCR $\downarrow$	GSS $\uparrow$	GSS m $\uparrow$	GSS h $\uparrow$
$K=1$	0.6356	0.2288	0.9979	0.9980	0.9973
$K=3$	0.6356	0.2288	0.9980	0.9981	0.9975
$K=5$	0.6610	0.2203	0.9979	0.9980	0.9972
$K=8$	0.6271	0.2373	0.9979	0.9981	0.9972
$K=10$	0.6525	0.2119	0.9979	0.9981	0.9972

Table 5: Effect of candidate number

K

in VLM verification. Bold : best result; underlined : second-best result.

Figure 4: Failure cases of CRGR caused by limited reference coverage, severe degradation, and glyph ambiguity.

Explore similar work

Think Before You Restore: Risk-Aware Manchu Manuscript Restoration with Stroke-Guided Attention
Sep 28, 2026Mingqiu Liang, Dongdong Wang, Siyang Lu +2Image Restoration Historical Document Processing
School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China · College of Design, Construction, and Planning, University of Florida, Gainesville, FL, USA · School of Japanese Studies, Dalian University of Foreign Languages, Dalian, China
DDVR-Net: Degradation-Aware Decoupling and Visibility-Guided Restoration for Historical Document Recognition
Oct 12, 2026Jiu Sha, Yu Weng, Mengxiao ZhuOptical Character Recognition Historical Document Processing
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
DocRevive: A Unified Pipeline for Document Text Restoration
Apr 11, 2026Kunal Purkayastha, Ayan Banerjee, Josep Llados +1Document Image Understanding
Computer Vision Center · Indian Statistical Institute, Kolkata

Sep 28, 2026J/K move · Enter open · S save

Think Before You Restore: Risk-Aware Manchu Manuscript Restoration with Stroke-Guided Attention

Mingqiu Liang, Dongdong Wang, Siyang Lu, Ting Huang+1

School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China · College of Design, Construction, and Planning, University of Florida, Gainesville, FL, USA · School of Japanese Studies, Dalian University of Foreign Languages, Dalian, China

Image Restoration Historical Document Processing

Open paper AskPDF arXiv

Full-page blind restoration of historical Manchu manuscripts is challenging due to scarce annotations, unknown degradation regions, and fragile connected strokes. Generic restoration models may improve visual quality but often modify intact content, leading to over-restoration. We propose SAGE-Restore (Stroke-Aware Gated rEstoration), a selective restoration framework that first assesses where restoration is needed and then uses this assessment to guide restoration candidate generation and pixel-level selection. Its encoder predicts patch-level repair probabilities from complementary appearance and stroke-structural cues to condition restoration candidate generation, while the corresponding repair logits are refined into a pixel-level soft gate that selectively controls where the restoration candidate is applied. We further introduce a fidelity-aware evaluation protocol that jointly measures degraded-region recovery, intact-content preservation, and their balance. SAGE-Restore achieves the highest R-Recovery (0.463) and RFS (0.626), while maintaining high U-Fidelity (0.968), demonstrating an effective balance between restoration and content preservation.

Abstract

Explore similar work

Think Before You Restore: Risk-Aware Manchu Manuscript Restoration with Stroke-Guided Attention

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

DocRevive: A Unified Pipeline for Document Text Restoration