Single-pixel imaging with untrained convolutional autoencoder network.

• We propose a physical model-driven untrained deep convolutional autoencoder network for SPI and validate its performance from simulations and experiments. • We designed an end-to-end SPI reconstruction network, which can better reconstruct high-quality images from under-sampled measurements. • We perform a comparative study through the simulations and experiments. The results demonstrate that UCAN outperforms other existed SPI methods, including DGI, TVAL3, and GIDC. Single-pixel imaging (SPI) is a novel imaging modality which captures the images with a single-pixel detector by using a lot of time-varying modulation patterns. Nowadays, SPI reconstructions with data-driven deep learning had been verified for high-quality reconstructions under low sampling ratios. However, it faces a dilemma of hard-to-get sufficient training sets in many practical applications, e.g., long-range single-pixel imaging fields. Here, a model-driven SPI reconstruction method based on untrained convolutional autoencoder network (UCAN) is proposed. This framework does not need to pre-train on any dataset and can be automatically optimized, then eventually produce the restored images through the interplay between the neural network and the SPI physical model. Simulations confirm the superiorities of the proposed method over many other existed algorithms in the SPI field. Also, the reconstructions for long-range single-pixel imaging in real urban atmospheric environments demonstrate that our method has better denoising performance. We believe that the present work provides an alternative framework for SPI and paves the way for practical applications, e.g., long-range optical remote sensing and low-irradiative biological imaging. [ABSTRACT FROM AUTHOR]