HSGAN: Hyperspectral Reconstruction From RGB Images With Generative Adversarial Network.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Hyperspectral (HS) reconstruction from RGB images is crucial for various applications.
• Current methods using convolutional neural networks lack consistent performance across diverse scenes and input image quality.
• Existing approaches struggle with real-world noisy RGB images.

Purpose of the Study:

• To enhance the accuracy and robustness of HS reconstruction from RGB images.
• To develop a framework that performs consistently across different scenes and input image types (clean and noisy).
• To address the limitations of current state-of-the-art HS reconstruction techniques.

Main Methods:

• Proposed an effective HSGAN framework utilizing a two-stage adversarial training strategy.
• Developed a generator with a four-level top-down architecture for multi-scale feature extraction and combination.
• Introduced a spatial-spectral attention block (SSAB) to capture spatial-wise and channel-wise relations for improved generalization to noisy images.

Main Results:

• HSGAN demonstrated superior performance in HS reconstruction compared to existing methods.
• Experiments were conducted on five well-known HS datasets using both clean and real-world noisy RGB images.
• The proposed SSAB effectively improved the model's ability to handle noisy input data.

Conclusions:

• The HSGAN framework offers a significant advancement in hyperspectral image reconstruction from RGB data.
• The two-stage adversarial training and SSAB contribute to improved accuracy and robustness.
• HSGAN provides a promising solution for reconstructing HS images from challenging real-world conditions.