Deep reconstruction methods for computational hyperspectral imaging
Title : Deep reconstruction methods for computational hyperspectral imaging
Asbtract : Deep learning is a very efficient framework to solve inverse problems in imaging. Following a recent trend, several neural network architectures have provided a link between deep and optimization-based image reconstruction methods. Contrary to the first “black box” networks, these novel deep-learning methods rely on a forward operator and lead to more interpretable results. In this presentation, we will describe the main network architectures for single-pixel image reconstruction. We will present reconstruction results from simulated and experimental acquisitions obtained with a spectral detector. We will show that deep neural networks can be trained easily, in an end-to-end manner, from simulation using databases such as STL-10 or ImageNet. We will finally show that the denoised completion network (DC-Net) and the expectation-maximization network (EM-Net) generalize very well to noise levels that are unseen during the training phase. However, the EM-Net is observed to achieve the same reconstruction quality as the DC-Net despite having fewer learnable parameters.
More information : https://www.creatis.insa-lyon.fr/~ducros/WebPage/index.html