.png){kind=icon}
.png){kind=icon}
Return
FPSR-GAN: fourier-guided progressive spatial refinement for high-fidelity thangka inpainting
.png){kind=icon}
PRE
AI
Y
Y
X
K
DOI：10.1007/s00530-026-02581-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Thangka images are characterized by complex line structures, repeated patterns, fine-grained textures, and distinctive artistic styles. Inpainting local damage in such images requires not only plausible completion of missing content, but also consistency with surrounding regions in structural direction, texture distribution, and stylistic expression, placing high demands on long-range structural modeling and local detail recovery. Although existing deep inpainting methods have achieved favorable results on natural images, they still struggle to simultaneously preserve global geometric consistency and local texture fidelity when handling large irregular missing regions and complex Thangka structures. Spatial-domain methods are limited by local receptive fields, which may lead to structural discontinuities and unnatural edge transitions, while directly introducing frequency-domain global modeling may cause spectral disturbances and spatial artifacts, affecting the naturalness and stability of restored results. To address these issues, this paper proposes a Fourier-Guided Progressive Spatial Refinement Generative Adversarial Network (FPSR-GAN), which integrates frequency-domain global modeling and spatial-domain multi-scale enhancement into an end-to-end inpainting framework. Specifically, the Frequency-Adaptive Global Learning (FAGL) module models global structural information in the frequency domain while suppressing artifacts, and the Spatial-Adaptive Multi-Scale (SAMS) module enhances multi-scale details and pixel-level selective focusing in the spatial domain. Their collaboration enables progressive feature refinement from global structural consistency to local texture fidelity. Moreover, a joint spectral-spatial constraint combines spatial-domain reconstruction, perceptual, style, and adversarial constraints with a spectral-domain focal frequency constraint, jointly optimizing visual consistency and spectral detail distribution. Experiments on CelebA-HQ, Paris StreetView, and a self-constructed Thangka dataset demonstrate that FPSR-GAN achieves favorable texture clarity and structural coherence, providing a feasible solution for digital cultural heritage image inpainting.
Keywords:
Thangka inpainting
GAN
Frequency-spatial collaborative modeling
Multi-scale feature enhancement
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers