Adaptive Feature Fusion-Enhanced Vision Transformer Autoencoder for Realistic Image Inpainting
Metaverse · 23.03.2026 · 7 (1)
DOI: 10.54517/m8247 ↗Kurz gefasst
Fachartikel in der Zeitschrift Metaverse, online erschienen am 23.03.2026. Verfasst von Jini P und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Image inpainting restores missing or corrupted image regions using contextual information from surrounding pixels. In the Metaverse integration of Virtual Reality (VR), Augmented Reality (AR), and Artificial Intelligence (AI)-realism and visual continuity are vital for immersive user experiences. This paper presents Vi-Trans, a Vision Transformer-based autoencoder for high-quality image inpainting. The model divides images into non-overlapping patches and reconstructs masked regions using global contextual learning through multi-head self-attention and feed-forward layers. To enhance structural integrity and edge preservation, a novel Adaptive Feature Fusion (AFF) module is introduced to fuse global transformer representations with local encoder features through an attention-weighted mechanism. This dynamic fusion balances semantic understanding with fine-grained spatial details, improving visual consistency. Experimental evaluations on the CelebA dataset demonstrate that Vi-Trans with AFF outperforms existing transformer-based inpainting methods in PSNR, SSIM, FID, and LPIPS metrics.
Zitieren
APA
P, J., & K.K., R. (2026). Adaptive Feature Fusion-Enhanced Vision Transformer Autoencoder for Realistic Image Inpainting. Metaverse, 7(1). https://doi.org/10.54517/m8247
Vancouver
P J, K.K. R. Adaptive Feature Fusion-Enhanced Vision Transformer Autoencoder for Realistic Image Inpainting. Metaverse. 2026;7(1). doi:10.54517/m8247
BibTeX
@article{P2026,
author = {P, Jini and K.K., Rajkumar},
title = {Adaptive Feature Fusion-Enhanced Vision Transformer Autoencoder for Realistic Image Inpainting},
journal = {Metaverse},
year = {2026},
volume = {7},
number = {1},
doi = {10.54517/m8247}
}
EUVIMED