A fast computational approach for high spectral resolution imaging

Abstract

Spectral imaging is at the cornerstone of many fields, including astronomy, environmental monitoring, food processing, agriculture, and biomedical imaging. While most current spectral techniques lack imaging speed, we describe a computational approach that allows for fast high spectral resolution imaging. Our set-up maps the image of the scene into the fiber of a compact spectrometer through a digital micromirror device (DMD), where a series of multiplexing patterns are uploaded. After a reconstruction step, the hypercube of the scene can be recovered. DMDs represent the fastest technology (e.g., >20 kHz) for spatial light modulators. The raw data is acquired for a (x,y,λ) hypercube of size 64x64x2048 in about 12 s, while the Hadamard inversion takes <1 s. The acquisition speed can be further reduced by limiting the number of Hadamard patterns; however, the resulting imaging reconstruction problem is turned into an underdetermined inverse problem, which requires regularization techniques to be used to obtain acceptable solutions, in particular, in the presence of noise. Deep learning is a very efficient framework to solve inverse problems in imaging. Following a recent trend, several convolution neural network architectures have provided a link between deep and optimization-based image reconstruction methods. Contrary to the initially proposed “black box” networks, these deep-learning methods rely on a forward operator and lead to more interpretable networks. Here, we review deep architectures for single-pixel image reconstruction and show that the network can be trained easily, in a end-to-end manner, using databases such as STL-10 or ImageNet. We present reconstruction results from simulated and experimental single-pixel acquisitions. We show that EM-Net generalizes very well to noise levels that are unseen during the training, despite having fewer learned parameters than alternative methods. The proposed EM-Net generally reconstructs images with fewer artifacts and with higher signal-to-noise ratios, particularly in high-noise situations.