Mid-infrared (MIR) imaging is crucial for diverse applications like biomedical diagnosis, industrial inspection, and remote sensing. However, current MIR imagers, primarily focal plane arrays (FPAs), face limitations such as high dark noise, low pixel counts, thermal susceptibility, high fabrication costs, and cryogenic requirements. The need for highly sensitive MIR imagers, particularly in low-photon-flux scenarios, motivates the search for alternative solutions. Single-pixel cameras, utilizing a single-element detector and spatial encoding masks, offer a cost-effective and robust alternative. While successful in visible and near-infrared regions, single-pixel MIR imaging at the single-photon level has remained elusive due to the lack of suitable detectors and spatial modulators. This paper addresses this challenge by presenting a novel approach to overcome the limitations of conventional MIR imaging technologies.

Existing MIR compressive imaging techniques primarily utilize few-pixel infrared sensors, falling short of single-photon sensitivity. Conventional spatial light modulators (SLMs), based on liquid crystals or micromirrors, have limited operational ranges and modulation accuracy in the MIR region. Although advancements in graphene metasurfaces show promise for MIR modulators, they are currently in the early stages of development. The combination of high-performance optical detectors and high-fidelity spatial modulators for single-photon MIR imaging has remained a significant hurdle. This work reviews prior efforts in single-pixel imaging (visible and near-infrared), including the use of Hadamard matrices and compressed sensing algorithms to reconstruct images from a limited number of measurements. The limitations of extending these techniques to the MIR range are discussed, highlighting the need for a novel approach.

The researchers developed a frequency-upconversion approach for single-pixel MIR imaging at the single-photon level. The core of their method is a nonlinear structured detection based on sum-frequency generation (SFG). A near-infrared structured pump beam, spatially modulated using a digital micromirror device (DMD), interacts with the MIR object image within a periodically poled lithium niobate (PPLN) crystal. The SFG process optically imprints the pump patterns onto the MIR radiation while simultaneously upconverting it into the visible region. A single-element silicon photodiode (Si-PD) detects the upconverted photons. The intensity measurements, correlated with the time-varying pump patterns, are used to reconstruct the MIR image. Hadamard matrices provide an orthonormal basis for efficient image reconstruction. For compressive imaging, random patterns and a primal-dual algorithm are employed. A deep convolutional neural network denoiser further improves image quality. Coincidence pulsed pumping and spectral-temporal optimization enhance sensitivity, enabling single-photon imaging with illumination intensities as low as 0.5 photons/pulse. Experiments were conducted using both analog silicon detectors and single-photon counting modules (SPCMs) based on silicon avalanche photodiodes. The experimental setup included synchronized dual-color ultrafast fiber lasers generating MIR (3070 nm) and near-infrared (1030 nm) pulses, a DMD for spatial modulation, a PPLN crystal for SFG, a spectral filter, and the detectors. Image reconstruction algorithms involved Hadamard transforms and compressed sensing techniques, along with a deep learning-based denoiser for improved performance in low-photon scenarios.

The researchers successfully demonstrated MIR single-pixel imaging at the single-photon level. They achieved real-time video recording at 10 Hz with 16x16 pixels and high-contrast images even at 0.5 photons/pulse illumination (with longer integration times). Compressive imaging with a 25% undersampling ratio was also demonstrated, achieving high-quality reconstructions using a deep learning denoiser, even at single-photon illumination levels. The system's performance was characterized using Hadamard patterns for standard single-pixel imaging and random patterns for compressive sensing. The use of Hadamard patterns enabled fast image reconstruction via matrix transposition, while random patterns allowed for sub-Nyquist sampling and subsequent image reconstruction using advanced compressed sensing and deep learning methods. Image quality was quantitatively assessed by comparing reconstructed images obtained from both simulated and experimental data. The system showed high sensitivity, enabling the detection of fine details in the test objects, such as the letters etched onto the surface of a metallic sheet or a silicon substrate. The improved signal-to-noise ratio is attributed to the combination of coincidence pulsed pumping for improved conversion efficiency, stringent spectral filtering for noise suppression, and the high sensitivity and speed of the single-photon detector. The spatial resolution was limited by the numerical aperture of the upconversion system and the phase-matching bandwidth of the nonlinear crystal, though suggestions for improvement are provided.

The successful demonstration of single-photon MIR single-pixel imaging addresses a critical gap in current MIR imaging capabilities. The nonlinear structured detection approach overcomes limitations posed by the lack of high-sensitivity MIR detectors and high-fidelity SLMs. The upconversion to the visible spectrum allows the use of highly sensitive and readily available silicon-based single-photon detectors at room temperature. The single-pixel configuration simplifies the system and reduces costs compared to FPA-based systems. The results showcase the potential of this technique for various applications requiring high sensitivity and low-light imaging in the MIR region, such as covert imaging, biomedical imaging and non-invasive chip defect inspection. The combination of compressed sensing and deep learning enhances the system's ability to reconstruct high-quality images even with limited photon counts and undersampling. Further improvements in spatial resolution and field of view can be achieved through optimization of the nonlinear crystal and potentially using advanced computational imaging techniques.

This work presents a groundbreaking advancement in MIR imaging by achieving single-photon sensitivity using a single-pixel architecture. The frequency-upconversion method successfully circumvents the limitations of current MIR detection technologies. The simplicity, sensitivity, and room-temperature operation of this system establish a promising platform for diverse applications requiring low-light, high-sensitivity MIR imaging. Future research could focus on enhancing spatial resolution, expanding the field of view, and exploring advanced algorithms for even more efficient data acquisition and reconstruction. The successful implementation of this approach suggests its applicability to other long-wavelength regions, such as terahertz frequencies.

The spatial resolution of the current system is limited by the numerical aperture of the upconversion imaging system and the phase-matching bandwidth of the nonlinear crystal. The field of view is determined by the pump beam size and the crystal aperture. The study primarily used binary sparse objects, but the authors suggest the method is adaptable to more complex scenarios. Further optimization of the imaging system and advanced algorithms could further improve image quality, especially at very low photon flux.