# Self-Regression Learning for Blind Hyperspectral Image Fusion Without Label

arXiv: https://arxiv.org/abs/2103.16806 · HTML: https://mlsanity.com/paper/2103.16806/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 7 of its 25 references that are in the map.
- First posted on arXiv 31 Mar 2021.

## Builds on (most cited first, 7 of 7)

- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [Spectral Normalization for Generative Adversarial Networks](https://mlsanity.com/paper/1802.05957/) (2018)
- [Invertible Residual Networks](https://mlsanity.com/paper/1811.00995/) (2018)
- [Visualizing Deep Convolutional Neural Networks Using Natural Pre-Images](https://mlsanity.com/paper/1512.02017/) (2015)
- [A convex formulation for hyperspectral image superresolution via subspace-based regularization](https://mlsanity.com/paper/1411.4005/) (2014)
- [Cross-Attention in Coupled Unmixing Nets for Unsupervised Hyperspectral Super-Resolution](https://mlsanity.com/paper/2007.05230/) (2020)
- [Bayesian Fusion of Multi-Band Images](https://mlsanity.com/paper/1307.5996/) (2013)

## Built on by (most cited first, 1 of 1)

- [A Survey on Deep Domain Adaptation and Tiny Object Detection Challenges, Techniques and Datasets](https://mlsanity.com/paper/2107.07927/) (2021)

## Abstract (arXiv)

Hyperspectral image fusion (HIF) is critical to a wide range of applications in remote sensing and many computer vision applications. Most traditional HIF methods assume that the observation model is predefined or known. However, in real applications, the observation model involved are often complicated and unknown, which leads to the serious performance drop of many advanced HIF methods. Also, deep learning methods can achieve outstanding performance, but they generally require a large number of image pairs for model training, which are difficult to obtain in realistic scenarios. Towards these issues, we proposed a self-regression learning method that alternatively reconstructs hyperspectral image (HSI) and estimate the observation model. In particular, we adopt an invertible neural network (INN) for restoring the HSI, and two fully-connected network (FCN) for estimating the observation model. Moreover, SoftMax nonlinearity is applied to the FCN for satisfying the non-negative, sparsity and equality constraints. Besides, we proposed a local consistency loss function to constrain the observation model by exploring domain specific knowledge. Finally, we proposed an angular loss function to improve spectral reconstruction accuracy. Extensive experiments on both synthetic and real-world dataset show that our model can outperform the state-of-the-art methods