Research on Restoration and Completion Algorithm of Cultural Relics Based on Generative Adversarial Network and Diffusion Model

A decision-level fusion framework was proposed to solve the problem that the texture generated by the generative adversarial network (GAN) is sharp, but the large-scale loss is easy to drift, and the complementarity of the diffusion model is coherent but the fine-grained texture is easy to over-smooth in the restoration of cultural relics. An asymmetric dual-stream generator with feature representation isolation is constructed, and a defect-aware routing controller is designed. The local edge density, fragment compactness, and neighborhood semantic gradient are taken as inputs, and the path is generated by pixelwise allocation through online preference mining. The routing weight modulates the cross-attention of the diffusion stream to the adversarial stream texture features by multiplicative gating, and reverse-infuses the low-level structure prior of the diffusion stream. An adaptive fusion operator guided by routing gradient is proposed to generate the bandwidth point by point according to the weighted gradient tensor and stretch it anisotropic along the main direction of the structure tensor. The three style anchoring constraints of stroke direction, pigment color moment and texture energy spectrum slope are introduced. On the hierarchical incomplete test set, the PSNR of the method reaches 28.94 dB and 25.31 dB in the transition interval from L2 to L3, which is 1.48 dB and 1.29 dB higher than the optimal shallow fusion baseline. The boundary LPIPS is reduced to 0.087, and the style distance is reduced to 0.108. Ablation experiments verify the independent contribution of routing mechanism and feature-level interaction and show that decision-level fusion can effectively realize the complementarity of the two streams under the capability constraints of the existing generation model.

Authors

Institutions

Publication Details

Journal
Journal of Discovery Core
Published
2026-11-19
DOI
https://doi.org/10.67541/jdc2613
Primary Topic
Generative Adversarial Networks and Image Synthesis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Research on Restoration and Completion Algorithm of Cultural Relics Based on Generative Adversarial Network and Diffusion Model

Min Ma
Journal of Discovery Core
Generative Adversarial Networks and Image Synthesis
article

Research on Restoration and Completion Algorithm of Cultural Relics Based on Generative Adversarial Network and Diffusion Model

Min Ma
article en

Abstract

A decision-level fusion framework was proposed to solve the problem that the texture generated by the generative adversarial network (GAN) is sharp, but the large-scale loss is easy to drift, and the complementarity of the diffusion model is coherent but the fine-grained texture is easy to over-smooth in the restoration of cultural relics. An asymmetric dual-stream generator with feature representation isolation is constructed, and a defect-aware routing controller is designed. The local edge density, fragment compactness, and neighborhood semantic gradient are taken as inputs, and the path is generated by pixelwise allocation through online preference mining. The routing weight modulates the cross-attention of the diffusion stream to the adversarial stream texture features by multiplicative gating, and reverse-infuses the low-level structure prior of the diffusion stream. An adaptive fusion operator guided by routing gradient is proposed to generate the bandwidth point by point according to the weighted gradient tensor and stretch it anisotropic along the main direction of the structure tensor. The three style anchoring constraints of stroke direction, pigment color moment and texture energy spectrum slope are introduced. On the hierarchical incomplete test set, the PSNR of the method reaches 28.94 dB and 25.31 dB in the transition interval from L2 to L3, which is 1.48 dB and 1.29 dB higher than the optimal shallow fusion baseline. The boundary LPIPS is reduced to 0.087, and the style distance is reduced to 0.108. Ablation experiments verify the independent contribution of routing mechanism and feature-level interaction and show that decision-level fusion can effectively realize the complementarity of the two streams under the capability constraints of the existing generation model.

Journal of Discovery CoreVol. 1(3)
Shunyi Hospital of Beijing Traditional Chinese Medicine Hospital (CN)
Generative Adversarial Networks and Image Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon