Contents 1. Introduction
Ultra-high-speed imaging is vital for observing microscopic and transient physical phenomena [1]. To date, traditional CCD or CMOS cameras have made the imaging speeds up to millions of frames per second (fps) possible [2]. However, these high-speed cameras are generally silicon-based and have high sensibility only in the visible wavelength spectrum. As an alternative, single-pixel imaging requires only one photodiode, providing a shortcut for real-time imaging with low cost in almost all electromagnetic bands [3], [4] where the pixelated sensors are expensive or unavailable. For example, the mature InGaAs detectors can promote cost-effective single-pixel imaging in the short-wave infrared spectral region, which can be applied to imaging through some scattering media [5]. However, in order to obtain one image, conventional single-pixel imaging requires sequentially modulation of spatial light with a lot of patterns, which makes it impossible for them to image with ultra-high speeds. Here, we propose a single-pixel imaging approach with the help of a neural network algorithm, with an ultrahigh imaging speed reaching 60.3 Mfps. The objects can be detected by a single fiber probe, which makes it particularly suit to fiber endoscopy for in vivo applications [6].
