Adaptive polarization-multiplexing foveated ghost imaging method based on deep learning

Vladimir S. Shumigai,Vladislav S. Tuchin,Polina E. Moreva,Arina M. Starceva,Azat O. Ismagilov,Boris A. Nasedkin,Anton N. Tcypkin
DOI: https://doi.org/10.1016/j.optcom.2024.131385
IF: 2.4
2024-12-13
Optics Communications
Abstract:The challenges of remote sensing and area mapping have been recently addressed by using single-pixel imaging providing flexibility in modification and high sensitivity. This paper proposes a method for image reconstruction using polarization-multiplexing foveated ghost imaging and deep learning. The method suggested involves detection of noisy images and subsequent generation of modified foveated illumination patterns. In our proof-of-concept experiment, we used two polarization channels with independent illumination patterns and the YOLOv8 algorithm to accelerate the data acquisition rate. The images were reconstructed using a compressive sensing algorithm that employs the Moore–Penrose pseudoinverse for norm minimization. Two techniques were considered for detecting objects in noisy images: (i) processing the test image using filters, and (ii) training the model on a noisy dataset, with images simulating those produced by ghost imaging. According to the results, the second technique was selected for its efficient and accurate object detection in noisy images, while the first one failed to correctly detect target objects and required time-consuming filter application for each test image, significantly increasing the system's processing time. Using modified patterns and two polarization channels has been found to increase the data acquisition rate by 4.6 times. This makes the proposed method promising for high-speed demanding applications including remote sensing and object positioning.
optics
What problem does this paper attempt to address?