Enhancement of spatial noise tolerance in single-pixel imaging with a feature-extended deep neural network

To compensate for time-fluctuating spatial noise and reconstruct an object image, a deep learning-based single-pixel imaging (SPI) system using a neural network consisting of five transposed convolutional layers and three convolutional layers has been developed. In the present study, we proposed a new image reconstruction method using deep learning with a feature-extended time-division pattern-learning (TDPL) network, which further increased the number of features in each layer to enhance the tolerance to time-fluctuating spatial noise. Simulations and experiments were performed to compare the performance of the proposed network with that of conventional methods, such as computational ghost imaging, Hadamard single-pixel imaging, deep convolutional auto-encoder network (DCAN), and TDPL network. We found that the image quality of the reconstructed image using the proposed method is superior to that of conventional methods in any environment with time-fluctuating spatial noise. For example, the quality of an object image reconstructed using the proposed method improved by − 0.037 and − 0.014 in a root-mean-square error and + 0.083 and + 0.005 in a structural similarity compared to that using the DCAN and TDPL network, respectively, under time-fluctuating spatial noise with a standard deviation of 0.5. Therefore, the proposed deep learning-based SPI system with a feature-extended TDPL network is expected to be applied to various imaging or observation in an environment where conditions are likely to change, such as astronomical observations, remote monitoring, and optical wireless communications.