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

Mingqiu Liang; Dongdong Wang; Siyang Lu; Ting Huang; Yingjun Qi

cs.CVSep 28, 2026

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

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

Organizations: 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

Abstract

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.

Figures & tables

	Full-reference (FR)										No-reference (NR)
Method	PSNR $\uparrow$	SSIM $\uparrow$	MAE $\downarrow$	MSE $\downarrow$	RMSE $\downarrow$	MS-S $\uparrow$	FSIM $\uparrow$	GMSD $\downarrow$	VIF $\uparrow$	NCC $\uparrow$	NIQE $\downarrow$	BRIS $\downarrow$	IL-N $\downarrow$	MAN $\uparrow$	MUSIQ $\uparrow$	CLIP $\uparrow$
Blind-Omni [ 10 ]	23.68	0.98	0.01	0.01	0.07	0.96	0.95	0.10	0.89	0.94	3.93	25.61	34.79	0.54	60.19	0.65
DiffBIR [ 5 ]	17.26	0.58	0.06	0.02	0.14	0.81	0.89	0.20	0.09	0.77	6.80	17.71	40.01	0.59	64.97	0.77
GSDM [ 26 ]	18.89	0.91	0.04	0.02	0.12	0.87	0.84	0.17	0.44	0.78	4.25	30.17	30.85	0.47	61.08	0.68

Table 1: Comparison of full-reference (FR) and no-reference (NR) image quality metrics.

Figure 1: Restoration performance across different methods, with red boxes indicating fidelity preservation and blue boxes indicating restoration quality.

Figure 2: Overview of the proposed SAGE-Restore framework. The SGA encoder predicts patch-level repair logits $L_{p}$ , whose probabilities $P$ condition restoration candidate generation. The logits are further refined into a pixel-level soft gate $G^{t}$ , which selectively combines the degraded input $I_{d}^{t}$ and restoration candidate $I_{r}^{t}$ . Restored overlapping tiles are combined using Hann-weighted blending.

Figure 3: Qualitative comparison on two Manchu manuscripts. GSDM recovers degraded strokes but introduces substantial changes to intact content, while Blind-Omni largely preserves the degraded input without recovering missing structures. DiffBIR and HYPIR improve overall appearance but fail to reconstruct key missing strokes. In contrast, SAGE-Restore selectively restores damaged structures while preserving intact handwriting and paper texture, achieving a better balance between recovery and fidelity.

	Fidelity-aware			Full-reference (FR)						No-reference (NR)
Method	R-Rec. $\uparrow$	U-Fid. $\uparrow$	RFS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	MAE $\downarrow$	FSIM $\uparrow$	GMSD $\downarrow$	VIF $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	CLIP $\uparrow$
Blind-Omni	0.033	0.993	0.063	23.677	0.977	0.010	0.954	0.101	0.889	3.932	60.188	0.653
DiffBIR	0.062	0.848	0.116	17.264	0.576	0.057	0.886	0.204	0.088	6.795	64.967	0.774
GSDM	0.420	0.819	0.555	18.888	0.913	0.040	0.837	0.171	0.442	4.250	61.080	0.683
HYPIR	0.047	0.889	0.089	19.399	0.807	0.038	0.923	0.156	0.170	5.867	60.762	0.819
SAGE-Restore	0.463	0.968	0.626	26.077	0.975	0.006	0.971	0.094	0.856	4.056	60.478	0.688

Table 2: Quantitative comparison with state-of-the-art restoration methods. Best results are in bold and second-best results are underlined.

SGA	Gate	R-Rec.	U-Fid.	RFS	PSNR	SSIM	MAE
–	–	0.322	0.881	0.472	21.270	0.873	0.028
✓	–	0.448	0.925	0.596	22.980	0.914	0.014
–	✓	0.405	0.965	0.571	24.780	0.944	0.009
✓	✓	0.463	0.968	0.626	26.077	0.975	0.006

Table 3: Ablation results for SGA and pixel-level gating.

Explore similar work

MESA: A Training-Free Multi-Exemplar Deep Framework for Restoring Ancient Inscription Textures
Apr 19, 2026Vasileios Toulatzis, Sofia Theodoridou, Ioannis FudosImage Restoration Historical Manuscripts
University of Ioannina, Greece
Reconstructing Historical Manuscripts through MSI: The Potential of Contrast in Assessing Image Quality and Legibility
Aug 13, 2026Anna BregerImage Quality Assessment Historical Manuscripts
Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge, UK
Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models
Jul 24, 2026Gabeen Kim, Kyeongpil KangHistory Language Modeling
Department of AI Convergence, Kangwon National University · Department of Computer Science and Engineering, Kangwon National University

Apr 19, 2026, cs.CVJ/K move · Enter open · S save

MESA: A Training-Free Multi-Exemplar Deep Framework for Restoring Ancient Inscription Textures

Vasileios Toulatzis, Sofia Theodoridou, Ioannis Fudos

University of Ioannina, Greece

Image Restoration Historical Manuscripts Photorealistic Style Transfer Generative Adversarial Network Art Multi--Task Learning

Open paper AskPDF arXiv

Ancient inscriptions frequently suffer missing or corrupted regions from fragmentation, erosion, or other damage, hindering reading, and analysis. We review prior image restoration methods and their applicability to inscription image recovery, then introduce MESA (Multi-Exemplar, Style-Aware) -an image-level restoration method that uses well-preserved exemplar inscriptions (from the same epigraphic monument, material, or similar letterforms) to guide reconstruction of damaged text. MESA encodes VGG19 convolutional features as Gram matrices to capture exemplar texture, style, and stroke structure; for each neural network layer it selects the exemplar minimizing Mean-Squared Displacement (MSD) to the damaged input. Layer-wise contribution weights are derived from Optical Character Recognition-estimated character widths in the exemplar set to bias filters toward scales matching letter geometry, and a training mask preserves intact regions so synthesis is restricted to damaged areas. We also summarize prior network architectures and exemplar and single-image synthesis, inpainting, and Generative Adversarial Network (GAN) approaches, highlighting limitations that MESA addresses. Comparative experiments demonstrate the advantages of MESA. Finally, we provide a practical roadmap for choosing restoration strategies given available exemplars and metadata.

Abstract

Explore similar work

MESA: A Training-Free Multi-Exemplar Deep Framework for Restoring Ancient Inscription Textures

Reconstructing Historical Manuscripts through MSI: The Potential of Contrast in Assessing Image Quality and Legibility

Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models