Accelerating computed tomographic imaging spectrometer reconstruction using a parallel algorithm exploiting spatial shift-invariance

Abstract

Computed Tomographic Imaging Spectrometers (CTIS) capture hyperspectral images in realtime. However, post processing the imagery can require enormous computational resources; thus, limiting its application to nonrealtime scenarios. To overcome these challenges we developed a highly parallelizable algorithm that exploits spatial shift-invariance. To demonstrate the versatility of our new algorithm, we developed implementations on a desktop and an embedded graphics processing unit (GPU). To our knowledge, our results show the fastest image reconstruction times reported.

The Application of Convolutional Neural Networks for Tomographic Reconstruction of Hyperspectral Images

“…The latter is assumed in this work. As 10-30 EM iterations are typically required [14], we carry out 20 iterations. Both the construction of H and reconstruction of f are implemented in MATLAB with the help of built-in sparse matrix manipulations.…”

“…1a we show an example of a CTIS image of a ColorChecker Classic Mini (Col-orChecker) captured with a custom-made CTIS camera, which utilizes the same optical system layout as White et al in Ref. [14]. The CTIS image consists of a central zerothorder undiffracted scene image and four surrounding first-order diffractions.…”

“…1c, respectively, and a true RGB image of the ColorChecker is shown in As CTIS images are compressed and not as easy to analyze as hyperspectral cubes, fast and precise real-time reconstruction of the hyperspectral cube from a CTIS image is an important but challenging goal. In practice, the common dimension of the hyperspectral cube often exceeds 100 × 100 × 100, resulting in long reconstruction times and mediocre accuracy for existing algorithms [14][15][16]. Therefore, we consider neural networks to circumvent the limitations of current reconstruction algorithms.…”

“…A CTIS image contains a central zeroth-order undiffracted image surrounded by four first-order diffracted images. Different algorithms[14][15][16] are used to convert multiplexed 2-D CTIS images into 3-D hyperspectral cubes, which are easier to visualize and analyze. A CTIS camera is, however, portable and cheaper than hyperspectral imagers such as pushbroom and has a wider range of applications.…”

The hybrid approach -- Convolutional Neural Networks and Expectation Maximization Algorithm -- for Tomographic Reconstruction of Hyperspectral Images

“…However, H is usually rectangular and thus non-invertible, preventing the direct computation of H −1 . As a result, iterative algorithms [19][20][21] have been proposed to reconstruct f . The reconstruction time is unfortunately quite long and the accuracy is mediocre, especially for large images or high numbers of spectral channels, which hinders practical applications.…”