Biological tissues, fog, clouds, and other diffusers create light scattering, preventing direct imaging due to complex disordered speckle patterns forming. There has been a lot of effort in establishing methods to recover the target images from disordered speckle patterns, particularly in biological fields. Current algorithms work on visible wavelength band images, such as speckle correlation imaging, transmission matrix measurement, and optical phase conjugation-based wave-front modulation.
Low-Resolution (LR) Blurring the Features of the Speckle Particles
The evolution of imaging algorithms has extended light's wavelength from visible range to mid-wave infrared, verifying the viability of imaging through the diffuser in the infrared band.
Near-infrared (NIR) band has an advantage over the visible band due to the larger optical memory effect range (OME). However, the degradation issues associated with NIR cameras are more prominent than visible detectors due to the limited pixel size.
Moreover, at longer wavelengths, the optical system's diffraction-limited resolution declines. Low resolution blurs the speckle particles' features or corrupts the speckle pattern's statistical aspects, increasing the imaging method's demands. Reconstructing the high-resolution (HR) target buried in the LR speckle pattern is a fundamental difficulty in diffuser imaging with NIR light.
Alternative Methods to Solve Problems Associated with Imaging
Several studies have addressed the issue of optical scattering in the past. For instance, deep learning (DL) is an emerging tool for computational imaging, providing solutions to problems associated with imaging through diffusers with visible light.
Similarly, a generative adversarial network (GAN) is recently utilized for imaging through diffusers, providing better performance in imaging. For example, imaging through dynamic diffusers has been achieved through classified reconstruction via the generative adversarial network. A similar generative adversarial network under low photon flux conditions recovered further hidden targets behind the dynamic diffusers.
Similarly, targets from unpaired images were reconstructed within small scattering point spread functions (PSF) using a cycle-generative adversarial network.
These researches focused on diffusers' generalization, denoising, and recovery with visible light. However, previous studies have not addressed the mapping issue between the deteriorated speckle pattern and the high-resolution target in the near-infrared scattering imaging system.
The Methods Used in the Study
This study introduces a physically-based learning technique that enhances the near-infrared scattering imaging system's resolution by supplementing the degraded information.
Derived from concepts used in super-resolution imaging, the information degraded by the detector is compensable. The redundancy of the raw speckle pattern allows reconstructing the target via a sub-speckle pattern with a single frame satisfying a specific sampling coefficient. Hence, the speckle pattern's redundancy and the degraded resolution mechanism provide the physical basis for supplementing and reconstructing high-resolution targets via deep learning.
The researchers proved the study concepts by visualizing the supplement information during the learning procedure and demonstrated that the methods aid the lost resolution from two degradation models. Moreover, it was verified that this method could reconstruct the high-resolution target regulating parameters that affect imaging resolution, such as the target size, the distance between the diffuser and camera, and the pupil diameter.
Reconstructing the HR Target from the LR Speckle Pattern
Reconstructing high-resolution images from low-resolution speckle pattern face two inverse problems, i.e. speckle pattern’s SR imaging through the diffuser and target recovery from speckle pattern.
The study has classified low-resolution speckle patterns into two categories based on information channels. The first category includes a sub-speckle pattern with the information channel being cut off, while the other category includes a down-sampled global information channel.
Due to the loss of speckle pattern's properties in LR, hybrid input-output and the error reduction (HIO-ER) cannot reconstruct the high-resolution image. On the other hand, a generative adversarial network is capable of image generation and achieves the reconstruction of high-resolution targets by exploiting the speckle pattern's properties.
The learning approach discussed in the study reconstructs the HR target by using the speckle's physical feature to replenish the information lost due to the resolution deterioration. In addition, the study presents SR imaging of the sub-speckle pattern to observe the incremental knowledge acquired during the learning process.
Conclusion
This approach effectively performs SR reconstruction using just one frame of the LR sub-speckle pattern, in contrast to earlier methods of SR imaging via diffusers. Hence, this approach can acquire the accurate reconstruction of the degradation model even when little information is used, allowing it to recover the HR target concealed in the LR speckle pattern.
The recovery ability of this technology not only makes the resolution requirement of the near-infrared scattering imaging system much more manageable but also makes the possibility for practical applications much more promising.
Reference
Qianqian Cheng, Lianfa Bai, Jing Han, Enlai Guo (2022) Super-resolution imaging through the diffuser in the near-infrared via physically-based learning. Optics and Lasers in Engineering. https://www.sciencedirect.com/science/article/abs/pii/S0143816622002391
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.