Single pixel imaging (SPI) is a novel technique capturing 2D images using a bucket detector with a high signal-to-noise ratio, wide spectrum range and low cost. Conventional SPI projects random illumination patterns to randomly and uniformly sample the entire scene’s information. Determined by Nyquist sampling theory, SPI needs either numerous projections or high computation cost to reconstruct the target scene, especially for high-resolution cases. To address this issue, we propose an efficient single pixel imaging technique (eSPI), which instead projects sinusoidal patterns for importance sampling of the target scene’s spatial spectrum in Fourier space. Specifically, utilizing the centrosymmetric conjugation and sparsity priors of natural images’ spatial spectra, eSPI sequentially projects two π 2 -phase-shifted sinusoidal patterns to obtain each Fourier coefficient in the most informative spatial frequency bands. eSPI can reduce requisite patterns by two orders of magnitude compared to conventional SPI, which helps a lot for fast and high-resolution SPI.

Correlation is widely used to reconstruct the object image in ghost imaging (GI). But it only offers a linear proportion of the signal-to-noise ratios (SNR) to the number of measurements. We develop a Gerchberg-Saxton-like technique for GI image reconstruction in this manuscript. The proposed technique takes the advantage of the integral property of the Fourier transform, and treat the captured data as constraints for image reconstruction. We numerically and experimentally demonstrate the technique, and observe a nonlinear growth of the SNR value with respect to the number of measurements in the simulation. The proposed technique provides a different perspective of image reconstruction of GI, and will be beneficial to further explore its potential.

Ghost imaging has rapidly developed for about two decades and attracted wide attention from different research fields. However, the practical applications of ghost imaging are still largely limited, by its low reconstruction quality and large required measurements. Inspired by the fact that the natural image patches usually exhibit simple structures, and these structures share common primitives, we propose a patch-primitive driven reconstruction approach to raise the quality of ghost imaging. Specifically, we resort to a statistical learning strategy by representing each image patch with sparse coefficients upon an over-complete dictionary. The dictionary is composed of various primitives learned from a large number of image patches from a natural image database. By introducing a linear mapping between non-overlapping image patches and the whole image, we incorporate the above local prior into the convex optimization framework of compressive ghost imaging. Experiments demonstrate that our method could obtain better reconstruction from the same amount of measurements, and thus reduce the number of requisite measurements for achieving satisfying imaging quality.

Emerging sensor designs increasingly rely on novel color filter arrays (CFAs) to sample the incident spectrum in unconventional ways. In particular, capturing a near-infrared (NIR) channel along with conventional RGB color is an exciting new imaging modality. RGB+NIR sensing has broad applications in computational photography, such as low-light denoising, it has applications in computer vision, such as facial recognition and tracking, and it paves the way towards low-cost single-sensor RGB and depth imaging using structured illumination. However, cost-effective commercial CFAs suffer from severe spectral crosstalk. This cross-talk represents a major challenge in highquality RGB+NIR imaging, rendering existing spatially multiplexed sensor designs impractical. In this work, we introduce a new approach to RGB+NIR image reconstruction using learned convolutional sparse priors. We demonstrate high-quality color and NIR imaging for challenging scenes, even including highfrequency structured NIR illumination. The effectiveness of the proposed method is validated on a large dataset of experimental captures and simulated benchmark results which demonstrate that this work achieves unprecedented reconstruction quality.

Limited by long acquisition time of 2D ghost imaging, current ghost imaging systems are so far inapplicable for dynamic scenes. However, it's been demonstrated that nature images are spatiotemporally redundant and the redundancy is scene dependent. Inspired by that, we propose a content-adaptive computational ghost imaging approach to achieve high reconstruction quality under a small number of measurements, and thus achieve ghost imaging of dynamic scenes. To utilize content-adaptive inter-frame redundancy, we put the reconstruction under an iterative reweighted optimization, with non-uniform weight computed from temporal-correlated frame sequences. The proposed approach can achieve dynamic imaging at 16fps with 64×64-pixel resolution.

Scene-dependent adaptive compressive sensing (CS) has been a long pursuing goal that has huge potential to significantly improve the performance of CS. However, with no access to the ground truth, how to design the scene-dependent adaptive strategy is still an open problem. In this paper, a restricted isometry property (RIP) condition-based error-clamping is proposed, which could directly predict the reconstruction error, i.e., the difference between the current-stage reconstructed image and the ground truth image, and adaptively allocate more samples to regions with larger reconstruction error at the next sampling stage. Furthermore, we propose a CS reconstruction network composed of Progressively inverse transform and Alternating Bi-directional Multi-grid Network, named PiABM-Net, that could efficiently utilize the multi-scale information for reconstructing the target image. The effectiveness of the proposed adaptive and cascaded CS method is demonstrated with extensive quantitative and qualitative experiments, compared with the state-of-the-art CS algorithms.