MQViT-OSR: geometric-aware vision transformer for omnidirectional super resolution with multiquadric interpolation
Omnidirectional images (ODIs) are capable of showing the entire 360° of real-world scenes used in immersive fields of surveillance, robotics, and virtual reality. Most ODIs are obtained at low resolutions, which cause blurred edges, geometric distortions. To handle these problems, MQViT-OSR is proposed that captures long-range spatial dependencies and efficiently reconstructs panoramic images with a high level of fidelity and resolution. The MQViT-OSR generator to capture worldwide spatial relationships and retain spherical geometric structure of ODIs. Multiquadric interpolation is used as pre-processing step, and relativistic average discriminator loss employed to make images more realistic. The VGG-19 perceptual loss guarantees high-level feature agreement between generated and real images. The ERP transformed dataset derived from DIV2 K enhances the quality of reconstruction and achieves WS-PSNR of 33.90 dB and WS-SSIM of 0.875, thus performing better than the state-of-the-art methods.
Authors
- Prathibha S Nair
- K. Maharajan (ORCID: https://orcid.org/0000-0003-4353-3133)
Institutions
- Kalasalingam Academy of Research and Education (IN)
Publication Details
- Journal
- The Imaging Science Journal
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1080/13682199.2026.2754668
- Primary Topic
- Advanced Image Processing Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00