Thangka super-resolution diffusion model based on discrete cosine transform domain padding upsampling and high-frequency focused attention
PLOS One · 2025
DOI: 10.1371/journal.pone.0332904 ↗Kurz gefasst
Fachartikel in der Zeitschrift PLOS One, 2025. Verfasst von Xin Chen und 5 weiteren Autor:innen. Beteiligte Einrichtung: Key Laboratory of Linguistic and Cultural Computing, Ministry of Education, Northwest Minzu University, Lanzhou, Gansu, China. Hauptthemen laut MEDLINE: Image Processing, Computer-Assisted, Paintings, Models, Theoretical. Die Arbeit nennt 53 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities.
Abstract (Originalsprache)
Thangka is a traditional Tibetan painting art form, possessing profound cultural significance and a unique artistic style. Image super-resolution technology, as an effective means of digital preservation and restoration, plays an important role in maintaining the integrity and heritage of Thangka art. However, existing image super-resolution methods cannot be used for Thangka images due to the following reasons: (1) Thangka images are large in size and rich in content, making it difficult for existing models to restore the original textures of degraded Thangka images. (2) Thangka images have intricate textures, so the high-resolution Thangka images reconstructed by existing methods, which perform well on objective metrics, perform poorly in terms of human visual perception. To overcome these challenges, a Frequency-Domain Enhanced Diffusion Super-Resolution (FDEDiff) is proposed, consisting of three parts: (1) a High-Frequency Focused Cross Attention Mechanism (HFC-Attention), which can separate high-frequency features of images to guide the attention mechanism, improving the reconstruction quality of high-frequency details in the diffusion model; (2) a DCT Domain Padding Upsampling (DCT-Upsampling), which performs upsampling in the Discrete Cosine Transform (DCT) domain and improves the reconstruction of dense line areas in Thangka images by fully utilizing global information; (3) for the first time, we construct a Thangka image super-resolution dataset, which contains 82,688 pairs of 512 × 512 images. Qualitative and quantitative experiments demonstrate that the proposed method achieves state-of-the-art performance on the Thangka dataset, attaining a LPIPS score of 0.0815 (lower indicates better perceptual quality) and showing a 20% improvement in perceptual quality over baseline methods. Although the proposed method requires longer inference time due to the iterative nature of diffusion models, this computational trade-off is justified by the critical need for artistic authenticity in cultural preservation applications. Dataset is available at https://github.com/cvlabdatasets/ThangkaDatasets .
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Autor:innen und Institutionen
- Key Laboratory of Linguistic and Cultural Computing, Ministry of Education, Northwest Minzu University, Lanzhou, Gansu, China.
- School of Mathematics and Computer Science, Northwest Minzu Univsersity, Lanzhou, China.
ORCID-Profile: Xin Chen, Nianyi Wang
Förderung
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Chen, X., Ji, L., Wang, Z., Yang, Y., Zhang, X., & Wang, N. (2025). Thangka super-resolution diffusion model based on discrete cosine transform domain padding upsampling and high-frequency focused attention. PLOS One. https://doi.org/10.1371/journal.pone.0332904
Vancouver
Chen X, Ji L, Wang Z, Yang Y, Zhang X, Wang N. Thangka super-resolution diffusion model based on discrete cosine transform domain padding upsampling and high-frequency focused attention. PLOS One. 2025. doi:10.1371/journal.pone.0332904
BibTeX
@article{Chen2025,
author = {Chen, Xin and Ji, Liqi and Wang, Zhen and Yang, Yunbo and Zhang, Xinyang and Wang, Nianyi},
title = {Thangka super-resolution diffusion model based on discrete cosine transform domain padding upsampling and high-frequency focused attention},
journal = {PLOS One},
year = {2025},
doi = {10.1371/journal.pone.0332904}
}
EUVIMED