PZRes-Net

PZRes-Net: Deep Neural Network for Hyperspectral Image Super-Resolution

PZRes-Net performs hyperspectral image (HSI) super-resolution by fusing low-resolution HSI (LR-HSI) with high-resolution multispectral images (HR-MSI) to reconstruct high-resolution HSI (HR-HSI) using spatial and spectral information across modalities.

Key Features:

• Wavelet-Inspired Fusion Framework: Implements a lightweight deep neural network architecture inspired by wavelet decomposition-based image fusion techniques to learn high-resolution residual representations.
• Spectral-Spatial Separable Convolution: Applies spectral-spatial separable convolution with dense connections to model cross-modality spatial and spectral dependencies.
• Zero-Mean Normalization: Enforces zero-centric residual learning through zero-mean normalization of feature maps at each network layer.
• Progressive Spectral Learning: Learns high-frequency spatial details progressively along the spectral dimension across all spectral bands.
• Mean-Value Invariant Residual Superimposition: Superimposes the learned high-resolution residual onto an up-sampled LR-HSI to generate a coarse HR-HSI prior to refinement.

Scientific Applications:

• Hyperspectral Image Reconstruction: Enhances spatial resolution of HSIs by integrating LR-HSI and HR-MSI data while preserving spectral fidelity.
• Cross-Modality Image Fusion: Resolves distribution discrepancies between hyperspectral and multispectral modalities for high-resolution reconstruction.

Methodology:

The network learns a zero-centric high-resolution residual image from LR-HSI and HR-MSI inputs using spectral-spatial separable convolutions with dense connections and zero-mean normalization. The residual is added to an up-sampled LR-HSI in a mean-value invariant manner to form a coarse HR-HSI, which is subsequently refined by jointly exploiting coherence across all spectral bands. Performance evaluation demonstrates improvements exceeding 3 dB in Peak Signal-to-Noise Ratio (PSNR) with 2.3× fewer parameters and 15× lower floating-point operations (FLOPs) compared to state-of-the-art methods.