A Comparative Evaluation of Super-Resolution Methods for Spectral Images Using Pretrained RGB Models
Year: 2026
Type: article
Abstract: The spatial resolution of spectral imaging systems is fundamentally constrained by hardware trade-offs, and the availability of large-scale annotated spectral datasets remains limited. This study presents a comprehensive evaluation of super-resolution (SR) methods across interpolation-based, CNN-based, GAN-based, and diffusion-based approaches. Using a synthetic 30-band spectral representation reconstructed from RGB with the MST++ model as a proxy ground truth, we arrange non-adjacent triplets as three-channel PNG inputs to ensure compatibility with existing SR architectures. A unified pipeline enables reproducible evaluation at ×2, ×4, and ×8 scales on 50 unseen images, with performance assessed using PSNR, SSIM, and SAM. Results confirm that bicubic interpolation remains a spectrally reliable baseline; shallow CNNs (SRCNN, FSRCNN) generalize well without fine-tuning; and ESRGAN improves spatial detail at the expense of spectral accuracy. Diffusion models (SR3, ResShift, SinSR), evaluated in a zero-shot setting without spectral-domain adaptation, exhibit unstable performance and require spectrum-aware training to preserve spectral structure effectively. The findings underscore a persistent trade-off between perceptual sharpness and spectral fidelity, highlighting the importance of domain-aware objectives when applying generative SR models to spectral data. This work provides reproducible baselines and a flexible evaluation framework to support future research in spectral image restoration.
Source: Sensors
DOI: 10.3390/s26020683
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
Funder: Université de Bourgogne
OpenAlex ID: W7124983470