|
Synthetic aperture techniques and concepts are widely used in radar, far-field imaging, and other important detection areas, where the application of synthetic aperture has stringent requirements for the quality and quantity of individual apertures, so there has been a significant challenge in the application of non-interferometric laser synthetic aperture techniques - reconstructing high-resolution images. The image information needs to be acquired for a long time to constrain the phase, and the results need to be reconstructed for a long time as well. In particular, for imaging rough surfaces, the removal of scattering noise often needs to be considered. Inspired by camera array imaging and deep learning in computational imaging, we uniquely apply camera arrays to optical synthetic aperture imaging systems and introduce deep learning to achieve high-resolution, high signal-to-noise imaging results of real far-field targets in a single shot. Specifically, we design a 3 × 3 camera array to acquire nine low-resolution images in parallel, each low-resolution image corresponds to a different region on the spectrum, while using a pairwise regression network to integrate the steps of sub-map acquisition, phase retrieval, and image denoising into a unified imaging system framework, driven by the data set, the neural network can quickly find the low-resolution, low-signal-to-noise ratio input in the system framework end to the high-resolution, high-signal-to-noise output end of the system framework, rather than the usual step-by-step solution. The results show that our system is able to improve the F-number from 12 to four equivalent imaging results in 0.2s, achieving a threefold increase in resolution, which allows for continuous high-resolution imaging results compared to traditional long acquisition and reconstruction. This method has the potential to break the limitations of current optical synthetic apertures and open up new avenues for fast frame rate high resolution imaging techniques.
|
