Deep-Learning Supervised Snapshot Compressive Imaging Enabled by an End-to-End Adaptive Neural Network

Authors:
Publisher website
Google Scholar
Abstract
Snapshot compressive imaging (SCI) is an advanced approach for single-shot high-dimensional data visualization. Deep learning is popularly used to improve SCI's performance. However, most existing methods are merely used as a replacement for analytical-modeling-based image reconstruction. Moreover, these models cling to the conventional random coded apertures and often presume a linear shearing operation. To overcome these limitations, we develop a new end-to-end convolutional neural network, termed deep high-dimensional adaptive net (D-HAN) that offers multi-faceted supervision to SCI by optimizing the coded aperture, sensing the shearing operation, and reconstructing three-dimensional datacubes. The D-HAN is implemented in two representative SCI systems for ultrahigh-speed imaging and hyperspectral imaging. The D-HAN is envisioned to benefit SCI in system design, image reconstruction, and performance evaluation.
Funding Information
  • Natural Sciences and Engineering Research Council of Canada (CRDPJ-532304-18, I2IPJ-555593-20, RGPAS-507845-2017, RGPIN-2017-05959, RTI-2019-00568)
  • Fonds de Recherche du Qubec - Sant (267406, 280229)
  • Emerging Leaders in the Americas Program ELAP
  • Canada Foundation for Innovation and Ministre de lconomie et de lInnovation du Qubec (37146)
  • Ministre des Relations internationales et de la Francophonie du Qubec
  • Fonds de Recherche du Qubec - Nature et Technologies (2019-NC-252960)
  • Sistema General de Regalas de Colombia (BPIN 2020000100415)

This publication has 82 references indexed in Scilit: