Spectral Super-Resolution via Deep Low-Rank Tensor Representation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Hyperspectral imaging (HSI) offers rich spectral information but acquiring high-resolution data is challenging.
• Current Convolutional Neural Network (CNN)-based spectral super-resolution methods often overlook the inherent low-rank property of HSIs, leading to high computational and storage demands.
• The limited receptive field of CNNs restricts their ability to capture global information correlations.

Purpose of the Study:

• To develop a novel network for spectral super-resolution that addresses the limitations of existing CNN-based approaches.
• To leverage the low-rank prior of hyperspectral images for efficient and effective spectral super-resolution.
• To reduce the computational and storage costs associated with hyperspectral image super-resolution.

Main Methods:

• A Low-Rank Tensor Reconstruction Network (LTRN) is proposed, treating HSI features as low-rank 3-D tensors.
• An Adaptive Low-Rank Prior Learning (ALPL) module combines Canonical-Polyadic (CP) decomposition with neural networks for 1-D feature learning.
• The ALPL module incorporates an Adaptive Vector Learning (AVL) module for HSI compression and a Multidimensionwise Multihead Self-Attention (MMSA) module to capture long-range dependencies.

Main Results:

• The LTRN effectively reconstructs hyperspectral images with enhanced spectral resolution.
• Experimental results on the CAVE and Harvard datasets demonstrate that the LTRN achieves comparable effectiveness to state-of-the-art methods.
• The proposed LTRN significantly reduces the number of parameters compared to existing approaches.

Conclusions:

• The LTRN provides a computationally efficient and effective solution for spectral super-resolution.
• The integration of low-rank tensor decomposition and self-attention mechanisms is beneficial for HSI reconstruction.
• The method offers a promising alternative for acquiring high-quality hyperspectral images with reduced resource requirements.