A Comparative Evaluation of Super-Resolution Methods for Spectral Images Using Pretrained RGB Models.

Area of Science:

• * Spectral imaging
• * Computer vision
• * Image processing

Background:

• * Hardware limitations and scarce annotated datasets constrain spectral imaging resolution.
• * Super-resolution (SR) techniques offer a potential solution for enhancing spatial detail in spectral data.
• * Evaluating diverse SR methods is crucial for advancing spectral image restoration.

Purpose of the Study:

• * To comprehensively evaluate interpolation-based, CNN-based, GAN-based, and diffusion-based SR methods for spectral imaging.
• * To establish a reproducible framework for assessing SR performance on spectral data.
• * To identify the trade-offs between spatial enhancement and spectral fidelity in SR models.

Main Methods:

• * Development of a synthetic 30-band spectral dataset using MST++ for ground truth.
• * Inputting non-adjacent RGB triplets into existing SR architectures for compatibility.
• * Evaluating SR models (bicubic, CNNs, ESRGAN, diffusion models) at ×2, ×4, and ×8 scales using PSNR, SSIM, and SAM metrics.

Main Results:

• * Bicubic interpolation serves as a spectrally reliable baseline.
• * Shallow CNNs demonstrate good generalization without fine-tuning.
• * ESRGAN enhances spatial detail but compromises spectral accuracy.
• * Diffusion models show unstable performance without spectral-domain adaptation, necessitating spectrum-aware training.

Conclusions:

• * A persistent trade-off exists between perceptual sharpness and spectral fidelity in SR.
• * Domain-aware objectives are critical for generative SR models applied to spectral data.
• * The study provides reproducible baselines and an evaluation framework for future spectral image restoration research.