Large depth-of-field (DOF) imaging with a high resolution is useful for applications ranging from robot vision to bio-imaging. However, it is challenging to construct an optical system with both a high resolution and large DOF. The common solution is to design relatively complex optical systems, but the setup of such systems is often bulky and expensive. In this paper, we propose a novel, compact, and low-cost method for large-DOF imaging. The core concept is to (1) design an aspherical lens with a depth-invariant point spread function to enable uniform image blurring over the whole depth range and (2) construct a deep learning network to reconstruct images with high fidelity computationally. The raw images captured by the aspherical lens are deblurred by the trained network, which enables large-DOF imaging at a smaller F number. Experimental results demonstrate that our end-to-end computational imager can achieve enhanced imaging performance. It can reduce loss by up to 46.5% compared to inherited raw images. With the capabilities of high-resolution and large-DOF imaging, the proposed method is promising for applications such as microscopic pathological diagnosis, virtual/augmented reality displays, and smartphone photography.

Single-pixel imaging (SPI) faces significant challenges in reconstructing high-quality images under complex real-world degradation conditions. This paper presents an innovative degradation model for the physical processes in SPI, providing the first comprehensive and quantitative analysis of various SPI noise sources encountered in real-world applications. Especially, pattern-dependent global noise propagation and object jitter modelling methods for SPI are proposed. Subsequently, a deep-blind neural network is developed to remove the necessity of obtaining parameters of all the degradation factors in real-world image compensation. Our method can operate without degradation parameters and significantly improve the resolution and fidelity of SPI image reconstruction. The deep-blind network training is guided by the proposed comprehensive SPI degradation model that describes real-world SPI impairments, enabling the network to generalize across a wide range of degradation combinations. The experiment validates its advanced performance in real-world SPI imaging at ultra-low sampling rates. The proposed method holds great potential for applications in remote sensing, biomedical imaging, and privacy-preserving surveillance.

Achieving real-time and high-accuracy 3D reconstruction of dynamic scenes is a fundamental challenge in many fields, including online monitoring, augmented reality, and so on. On one hand, traditional methods, such as Fourier transform profilometry (FTP) and phase-shifting profilometry (PSP), are struggling to balance measuring efficiency and accuracy. On the other hand, deep learning-based approaches, which offer the potential for improved accuracy, are hindered by large parameter amounts and complex structures less amenable to real-time requirements. To solve this problem, we proposed a network architecture search (NAS)-based method for real-time processing and 3D measurement of dynamic scenes with rate equivalent to single-shot. A NAS-optimized lightweight neural network was designed for efficient phase demodulation, while an improved dual-frequency strategy was employed coordinately for flexible absolute phase unwrapping. The experiment results demonstrate that our method can effectively perform 3D reconstruction with a reconstruction speed of 58fps, and realize high-accuracy measurement of dynamic scenes based on deep learning for what we believe to be the first time with the average RMS error of about 0.08 mm.

High-quality imaging with reduced optical complexity has been extensively investigated owing to its promising future in academic and industrial research. However, the practical performance of most imaging systems has encountered a bottleneck posed by optics rather than electronics. Here, we propose a digital lens (DL) to compensate for the chromatic aberration induced by physical optical elements, while the residual wavelength-independent degradation is tackled through a self-designed neural network. By transforming physical aberration correction to an algorithm-based computational imaging task, the proposed DL enables our framework to reduce optical complexity and achieve achromatic imaging in the analog domain. Real experiments have been conducted with an off-the-shelf single lens and recovered images show up to 14.62 dB higher peak signal-to-noise ratio (PSNR) than the original chromatic input. Furthermore, we run a comprehensive ablation study to highlight the contribution of embedding the proposed DL, which shows a 4.83 dB PSNR improvement compared with the methods without DL. Technically, the proposed method can be an alternative for future applications that require both simple optics and high-fidelity visualization.

Hyperspectral imaging is being extensively investigated owing to its promising future in critical applications such as medical diagnostics, sensing, and surveillance. However, current techniques are complex with multiple alignment-sensitive components and spatiospectral parameters predetermined by manufacturers. In this paper, we demonstrate an end-to-end snapshot hyperspectral imaging technique and build a physics-informed dual attention neural network with multimodal learning. By modeling the 3D spectral cube reconstruction procedure and solving that compressive-imaging inverse problem, the hyperspectral volume can be directly recovered from only one scene RGB image. Spectra features and camera spectral sensitivity are jointly leveraged to retrieve the multiplexed spatiospectral correlations and realize hyperspectral imaging. With the help of integrated attention mechanism, useful information supplied by disparate modal components is adaptively learned and aggregated to make our network flexible for variable imaging systems. Results show that the proposed method is ultra-faster than the traditional scanning method, and 3.4 times more precise than the existing hyperspectral imaging convolutional neural network. We provide theory for network design, demonstrate training process, and present experimental results with high accuracy. Without bulky benchtop setups and strict experimental limitations, this simple and effective method offers great potential for future spectral imaging applications such as pathological digital stain, computational imaging and virtual/augmented reality display, etc.