I. Introduction
Hyperspectral sensors record the reflectance or transmittance values of hundreds or thousands of bands ranging from infrared spectrum to ultraviolet spectrum. Unlike traditional RGB images, each pixel in a hyperspectral image (HSI) contains a continuous spectrum with abundant spectral signatures. Also, the abundance of spectral signatures facilitates in many computer vision fields, such as object tracking [1], image classification [2], [3], scene segmentation [4], and hyperspectral band selection [5]. Nevertheless, most existing hyperspectral devices utilize 2-D sensors to capture 3-D data through either spatial or spectral scanning, which need more exposure time and prevent their application to dynamic scenes [6]. In addition, the hyperspectral systems with heavy cost require a tradeoff between spectral resolution and spatial or temporal resolution [7]. In order to deal with the abovementioned problem, scanning-free or snapshot hyperspectral devices have been developed in the past decade, for instance, computed tomography imaging spectrometers (CTISs) [8], [9], hybrid RGB-HS systems [10], and aperture masks [11]–[13]. Unfortunately, these acquisition systems still require complex hardware devices and the spatial resolution of the acquired HSIs is still limited.