Double-Gated Mamba Multi-Scale Adaptive Feature Learning Network for Unsupervised Single RGB Image Hyperspectral...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Reconstructing hyperspectral images (HSI) from RGB images is challenging due to limited labeled data.
• Existing methods suffer from detail loss, poor robustness, and difficulty balancing spatial-spectral resolution.

Purpose of the Study:

• To develop an advanced network model for accurate and robust unsupervised hyperspectral image reconstruction.
• To address the limitations of current methods in detail preservation and spatial-spectral trade-off.

Main Methods:

• Proposed the Double-Gated Mamba Multi-Scale Adaptive Feature (DMMAF) learning network.
• Introduced adaptive dual-noise-aware feature extraction for edge details and robustness.
• Implemented deformable attention for global features and Mamba for local features to enhance information interaction.
• Developed a structure-aware smooth loss function to balance spatial-spectral resolution.

Main Results:

• Achieved state-of-the-art unsupervised reconstruction performance on NTIRE 2020, Harvard, and CAVE datasets.
• Demonstrated superior results compared to existing advanced algorithms.
• Attained specific performance metrics: NTIRE 2020 (MRAE 0.133, RMSE 0.040, PSNR 31.314), Harvard (RMSE 0.025, PSNR 34.955), CAVE (RMSE 0.041, PSNR 30.983).

Conclusions:

• The DMMAF network effectively overcomes limitations in unsupervised hyperspectral image reconstruction.
• The proposed methods significantly improve reconstruction accuracy, robustness, and spatial-spectral balance.
• The model shows strong potential for real-world applications requiring high-fidelity hyperspectral imaging.