High-performance deep-learning based polarization computational ghost imaging with random patterns and orthonormalization

Abstract

Abstract Polarization computational ghost imaging (PCGI) often requires a large number of samples to reconstruct the targets, which can be optimized by reducing sampling rates with the aids of deep-learning technology. In this paper, the random patterns and successive orthonormalization instead of common Hadamard patterns, has been introduced into the deep-learning based PCGI system to recover high-quality images at lower sampling rates. Firstly, we use a polarized light to illuminate the target with random patterns for sampling. Then we can obtain a vector of bucket detector values containing the reflective information of the target. Secondly, we orthonormalize the vector according to the random patterns. Subsequently, the orthonormalized data can be input into the Improved U-net (IU-net) for reconstructing the targets. We demonstrate that higher-quality image of the testing sample can be obtained at a lower sampling rate of 1.5%, and superior-generalization ability for the untrained complex targets can be also achieved at a lower sampling rate of 6%. Meanwhile, we have also investigated the generalization ability of the system for the untrained targets with different materials that have different depolarization properties, and the system still demonstrates superior performances. The proposed method may pave a way towards the real applications of the PCGI.