Implementation of hyperspectral inversion algorithms on FPGA: Hardware comparison using High Level Synthesis

Hyperspectral imaging is gathering significant attention due to its potential in various domains such as geology, agriculture, ecology, and surveillance. However, the associated processing algorithms, which are essential for enhancing output quality and extracting relevant information, are often computationally intensive and have to deal with substantial data volumes. Our focus lies on reconfigurable hardware, particularly recent FPGAs. While FPGA design can be complex, High Level Synthesis (HLS) workflows have emerged as a solution, abstracting low-level design intricacies and enhancing productivity. Despite successful prior efforts using HLS for hyperspectral imaging acceleration, we lack a comprehensive research to benchmark various algorithms and architectures within a unified framework. This study aims to quantitatively evaluate performance across different inversion algorithms and design architectures, providing insights for optimal trade-offs for specific applications. We apply this analysis to the case study of spectrum reconstruction processed from interferometric acquisitions taken by Fourier transform spectrometers.

Spectro-spatial hyperspectral image reconstruction from interferometric acquisitions

In the last decade, novel hyperspectral cameras have been developed with particularly desirable characteristics of compactness and short acquisition time, retaining their potential to obtain spectral/spatial resolution competitive with respect to traditional cameras. However, a computational effort is required to recover an interpretable data cube. In this work we focus our attention on imaging spectrometers based on interferometry, for which the raw acquisition is an image whose spectral component is expressed as an interferogram. Previous works have focused on the inversion of such acquisition on a pixel-by-pixel basis within a Bayesian framework, leaving behind critical information on the spatial structure of the image data cube. In this work, we address this problem by integrating a spatial regularization for image reconstruction, showing that the combination of spectral and spatial regularizers leads to enhanced performances with respect to the pixelwise case. We compare our results with Plug-and-Play techniques, as its strategy to inject a set of denoisers from the literature can be implemented seamlessly with our physics-based formulation of the optimization problem.