| Title |
Spectral Masked Diffusion Transformer for Synthetic Aperture Radar to Optical Image Translation |
| Authors |
윤선재(Sunjae Yoon) ; 홍지우(Ji Woo Hong) ; 구관형(Gwanhyeong Koo) ; 김규원(Kyuwon Kim) ; 안종식(JongSik Ahn) ; 이태영(Tae-Young Lee) ; 유창동(Chang D. Yoo) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.11.79 |
| Keywords |
Diffusion model; Synthetic aperture radar; Spectral masking; Transformer; Image translation |
| Abstract |
Synthetic Aperture Radar (SAR) imagery ensures reliable environmental and temporal coverage (e.g., through clouds and day-night cycles), but its noise and unique structural patterns make interpretation challenging, especially for non-experts. Translation from SAR to Optical (SAR-to-opt) image has emerged to make SAR images more perceptually interpretable. However, conventional generative methods rely on pixel-wise naive reconstruction, lacking a contextual understanding of the generated content. To this end, this paper proposes Spectral Masked Diffusion Transformer (SMDT) which performs image translation based on geographic context understanding by spectral masked reconstruction of optical images. The spectral masked reconstruction performs low-frequency component reconstruction, as SAR images contain rich contextual information in low frequencies. Experimental results on three retrieval benchmarks (i.e., QXS-SAROPT, SAR2Opt, SpaceNet6) achieve state-of-the-art performances and demonstrate the effectiveness of our SMDT. |