Single photon counting compressive imaging based on a sampling and reconstruction integrated deep network

Abstract

Single photon counting compressive imaging, a very efficient implementation of compressed sensing theory in photon counting imaging, offers the advantages of low cost and ultra-high sensitivity. However, when performing high-resolution imaging, single photon counting compressive imaging needs a long imaging time due to a lot of measurements and a large amount of image reconstruction calculations. In this paper, we demonstrate a single-photon counting compressed imaging system based a novel sampling and reconstruction integrated deep network. We call this network BF2C-Net. A binarized fully-connected layer is specially designed as the first layer of the network and trained out as a binary measurement matrix that can be directly loaded on the DMD to perform efficiently compressive sampling. The remaining layers of the network except the first layer are used to quickly reconstruct the compressed sensing image. The effects of compression sampling rate, measurement matrix and reconstruction algorithm on imaging performance are compared. The experimental results show that the BF2C-Net significantly outperforms existing iterative method and most other deep learning-based methods.