Hyperspectral Image Super-Resolution via Deep Progressive Zero-Centric Residual Learning.

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning

Background:

• Hyperspectral image (HSI) super-resolution is challenging due to cross-modality spatial and spectral information.
• Existing methods struggle with efficient fusion of low-resolution HSI and high-resolution multispectral images (HR-MSI).

Purpose of the Study:

• To develop an efficient and effective deep neural network for HSI super-resolution.
• To address the cross-modality fusion problem by learning a high-resolution residual image.

Main Methods:

• Proposed a novel lightweight deep neural network: progressive zero-centric residual network (PZRes-Net).
• Employed spectral-spatial separable convolution with dense connections for efficient residual learning.
• Introduced zero-mean normalization to achieve a zero-mean residual image.

Main Results:

• PZRes-Net significantly outperforms state-of-the-art methods on benchmark datasets.
• Achieved over 3dB improvement in Peak Signal-to-Noise Ratio (PSNR).
• Reduced parameters by 2.3x and Floating Point Operations (FLOPs) by 15x.

Conclusions:

• PZRes-Net offers a superior solution for hyperspectral image super-resolution.
• The model demonstrates high efficiency in terms of parameters and computation.
• Publicly available code facilitates further research and application.