Photon-limited non-imaging object detection and classification based on single-pixel imaging system

Under photon-limited detection which is limited by the low-light illumination and short detection time, off-the-shelf classification methods based on clear imaging of the object cannot achieve considerable classification accuracy. To solve this problem, we propose a non-imaging classification method based on single-pixel imaging system. With low-intensity pulsed illumination and time-correlated single-photon counting detection, binarized feature sequence of the objects that need to be classified can be obtained. Combining with a simple machine learning algorithm trained with simulated data based on Poissonian photon detection algorithm, the objects could be classified with considerable accuracy. Proof-of-principle experiments use the MNIST handwriting digit database, showing that up to 90% classification accuracy could be achieved with fewer than 1 detected photon per pixel.