GAN-SRSPI: super-resolution single-pixel imaging using generative adversarial networks

Single-Pixel Imaging (SPI) techniques enable the reconstruction of an image scene utilizing multiple spatially modulated light patterns and the corresponding measurements from a single-pixel detector. Since in SPI, the acquisition time scales quadratically with the image resolution. A high-resolution image reconstruction suffers from a slow reconstruction speed. This work proposes a super-resolution single-pixel imaging methodology based on generative adversarial networks (GAN-SRSPI). A low-resolution (N×N) image is reconstructed and then super-resolved to obtain a high-resolution (4N× 4N) image. The 4×super-resolution leads to an overall sampling rate of 6.25% (1/16). The previous work only minimizes the mean squared error on the pixel level. The reconstructed image fidelity of the high-frequency part needs to be improved. For the first time, a perceptual loss is proposed in the field of SPI super-resolution. The perceptual loss describes the perceptual similarity instead of pixel-level similarity. An adversarial loss differentiates the original high-resolution image from the super-resolved image. By combining the two, our results are more natural and more consistent with the perceptual characteristics of human eyes. The superiority of the proposed method over the traditional interpolation methods was visually demonstrated in the experiments. Our simulation shows a peak signal-to-noise ratio of 27.65 dB and structural similarity of 0.8076. Our work shows that GAN-SRSPI is a flexible and effective solution for high-resolution and fast SPI.

[1]  A. Forbes,et al.  Super-resolved quantum ghost imaging , 2022, Scientific Reports.

[2]  Zibang Zhang,et al.  Playing Tic-Tac-Toe Games with Intelligent Single-pixel Imaging , 2022, ArXiv.

[3]  Jie Cao,et al.  Ghost imaging through scattering medium by utilizing scattered light. , 2022, Optics express.

[4]  Graham M Gibson,et al.  Single-pixel imaging 12 years on: a review. , 2020, Optics express.

[5]  Fei Wang,et al.  Learning from simulation: An end-to-end deep-learning approach for computational ghost imaging. , 2019, Optics express.

[6]  Shensheng Han,et al.  Single-frame wide-field nanoscopy based on ghost imaging via sparsity constraints , 2019, Optica.

[7]  Takuo Tanemura,et al.  Ghost imaging using a large-scale silicon photonic phased array chip. , 2019, Optics express.

[8]  Miles J. Padgett,et al.  Principles and prospects for single-pixel imaging , 2018, Nature Photonics.

[9]  Roderick Murray-Smith,et al.  Deep learning for real-time single-pixel video , 2018, Scientific Reports.

[10]  Wen Chen,et al.  1000 fps computational ghost imaging using LED-based structured illumination. , 2018, Optics express.

[11]  Zhuo Xu,et al.  Underwater computational ghost imaging. , 2017, Optics express.

[12]  Christian Ledig,et al.  Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network , 2016, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[13]  Wenlin Gong,et al.  Three-dimensional ghost imaging lidar via sparsity constraint , 2016, Scientific Reports.

[14]  Andrew Zisserman,et al.  Very Deep Convolutional Networks for Large-Scale Image Recognition , 2014, ICLR.

[15]  Aline Roumy,et al.  Low-Complexity Single-Image Super-Resolution based on Nonnegative Neighbor Embedding , 2012, BMVC.

[16]  A. Gatti,et al.  Differential ghost imaging. , 2010, Physical review letters.

[17]  Ting Sun,et al.  Single-pixel imaging via compressive sampling , 2008, IEEE Signal Process. Mag..

[18]  Hanbo Wang.  Compressed Sensing: Theory and Applications , 2023, Journal of Physics: Conference Series.

[19]  Christopher K. I. Williams,et al.  International Journal of Computer Vision manuscript No. (will be inserted by the editor) The PASCAL Visual Object Classes (VOC) Challenge , 2022 .