Abstract
S2CycleDiff is a novel diffusion-based framework designed to overcome the inherent spatial-spectral trade-off in hyperspectral imaging. By employing a bilateral cycle-diffusion process, the model iteratively refines spatial and spectral details through mutual guidance. This approach leverages the generative power of Diffusion Probabilistic Models (DPM) to achieve high-fidelity reconstruction. Experimental evaluations on benchmark datasets like CAVE, Pavia Center, and Chikusei demonstrate that S2CycleDiff significantly outperforms existing state-of-the-art methods in both spatial detail preservation and spectral accuracy.
Introduction
Hyperspectral images (HSI) are critical for applications requiring precise material identification due to their high spectral resolution. However, physical sensor limitations often force a trade-off between spectral and spatial resolution. Hyperspectral Image Super-resolution (HISR) aims to mitigate this by fusing a low-resolution HSI (LrHSI) with a high-resolution multispectral image (HrMSI).
Traditional deep learning approaches often struggle with this fusion, leading to spectral distortion or spatial blurring. These models typically attempt a direct one-shot mapping that fails to fully exploit the complementary information between the two input sources. S2CycleDiff addresses these challenges by reframing HISR as an iterative, guided denoising process.
The S2CycleDiff Framework
The core of S2CycleDiff is the Bilateral Cycle-Diffusion process. Unlike standard models, it runs two parallel, interlinked paths: one focused on spatial enhancement (conditioned on LrHSI and guided by HrMSI) and another on spectral enhancement (conditioned on HrMSI and guided by LrHSI).
Key Components:
- Spatial-Spectral Guided Pyramid Denoising (SGPD): This module serves as the primary denoising engine. It utilizes a multi-scale pyramid structure equipped with Detail Injection Blocks (DIB). These blocks employ a cross-attention mechanism to intelligently extract and inject fine details from the guide images into the noisy hyperspectral data.
- Bilateral Guidance: The spatial path (Spa-GPD) and spectral path (Spe-GPD) operate in a cycle, ensuring that improvements in one domain inform and constrain the other. This prevents the "tug-of-war" common in fusion tasks where sharpening edges might otherwise degrade color signatures.
- Complementary Fusion Block: A final integration stage that merges the outputs of the bilateral processes to produce the final high-resolution HSI (HrHSI).
- Consistency Constraints: The framework is optimized using four distinct loss functions that enforce mathematical consistency between the generated output and the original input sensors via Point Spread Function (PSF) and Spectral Response Function (SRF) degradations.
Experimental Results
The researchers evaluated S2CycleDiff on three major datasets: CAVE (indoor), Pavia Center (urban), and Chikusei (aerial). Following Wald’s protocol, the model was tested against several state-of-the-art methods, including PSRT and SSR-NET.
S2CycleDiff achieved superior quantitative results across all metrics. For instance, on the CAVE dataset, it reached a PSNR of 43.92, significantly higher than the 43.47 achieved by the next-best method, PSRT. More importantly, it achieved the lowest Spectral Angle Mapper (SAM) scores, indicating that the reconstructed spectral signatures are highly accurate and reliable for downstream analysis.
Visual analysis through error maps confirms these findings. S2CycleDiff exhibits the lowest residuals in high-texture areas, maintaining sharp edges and clean spectral profiles even at high upsampling factors (e.g., x16) where other methods typically produce significant artifacts or blurring.
Conclusion
S2CycleDiff represents a significant advancement in HISR by successfully applying diffusion models to a bilateral fusion task. While the iterative nature of diffusion models introduces higher computational costs compared to one-shot CNNs, the gains in fidelity and spectral integrity are substantial. Future work will focus on accelerating the sampling process and adapting the framework to handle complex, real-world sensor noise and atmospheric interference.
Summary updated