Deep Learning-Based Hyperspectral Imaging

Case ID:

This technology is a pairwise-image-based hyperspectral convolution neural network (pHSCNN) that works to recover hyperspectral images from a pair of RGB images to improve the fidelity of hyperspectral images. The technology captures sequentially via a color sensor paired with and without an optimized optical filter in front of the imaging lens. This filter also works to improve overall system performance.

The technology proposed can be applied to biomedical, remote sensing, and computer vision applications to improve the resolution of hyperspectral images collected. Fidelity is a measure of how well two images may be distinguished, and so this technology could be of great use in aforementioned applications as spectral and spatial resolutions are improved via this technique.

Conventional hyperspectral imaging techniques currently suffer poor temporal, spatial, or spectral resolutions. However, computational hyperspectral recovery via RGB images has shown promise in that dynamic results can be delivered without sacrificing spectral or spatial resolutions. Despite the promise of this technology, performance still suffers at short and long wavelengths. To improve red and blue bands, the pairwise-image-based hyperspectral convolution neural network is proposed. Tested with a dual-camera hyperspectral system, pHSCNN is trained to reconstruct the hyperspectral images via RGB pairs and optimize the system as a whole. In the technology’s testing it is important to note that commercial filters were used and thus this saved cost on the manufacturing of the optimized filter and allows it to be realized in applications more readily.  


  • Biomedical imaging
  • Computer vision
  • Remote sensing
  • Hyperspectral imaging


  • Optimized system
  • Inexpensive
Patent Information:
Contact For More Information:
Richard Weite
Senior Licensing Manager, College of Optical Sciences
The University of Arizona
Lead Inventor(s):
Rongguang Liang
Yuanyuan Sun