EFFICIENT SPARSITY-BASED INVERSION FOR PHOTON-SIEVE SPECTRAL IMAGERS WITH TRANSFORM LEARNING


Kamaci U., Akyon F. C., Alkanat T., Oktem F. S.

5th IEEE Global Conference on Signal and Information Processing (GlobalSIP), Montreal, Canada, 14 - 16 November 2017, pp.1225-1229 identifier

  • Publication Type: Conference Paper / Full Text
  • Volume:
  • City: Montreal
  • Country: Canada
  • Page Numbers: pp.1225-1229
  • Middle East Technical University Affiliated: Yes

Abstract

We develop an efficient and adaptive sparse reconstruction approach for the recovery of spectral images from the measurements of a photon-sieve spectral imager (PSSI). PSSI is a computational imaging technique that enables higher resolution than conventional spectral imagers. Each measurement in PSSI is a superposition of the blurred spectral images; hence, the inverse problem can be viewed as a type of multi-frame deconvolution problem involving multiple objects. The transform learning-based approach reconstructs the spectral images from these superimposed measurements while simultaneously learning a sparsifying transform. This is performed using a block coordinate descent algorithm with efficient update steps. The performance is illustrated for a variety of measurement settings in solar spectral imaging. Compared to approaches with fixed sparsifying transforms, the approach is capable of efficiently reconstructing spectral images with improved reconstruction quality.