Adaptive locating foveated ghost imaging based on affine transformation

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Ghost imaging (GI) offers advantages like broad spectrum and anti-interference for applications such as spectral and 3D imaging.
• Limited sampling efficiency in traditional GI hinders its widespread adoption and real-time capabilities.

Purpose of the Study:

• To develop a novel deep learning-based method for enhancing the sampling efficiency of ghost imaging.
• To enable adaptive region of interest (ROI) selection for improved ghost imaging performance.

Main Methods:

• Proposed a foveated pattern affine transformer method utilizing a retina affine transformer (RAT) network for single-target ROI detection.
• Developed a recurrent neural network integrated with RAT (RNN-RAT) for multi-target ROI detection.
• Employed variable foveated speckle patterns and affine matrix prediction for adaptive ROI targeting.

Main Results:

• Achieved ROI localization and pattern generation in 0.358 ms, representing a 1x10^5 efficiency improvement over previous methods.
• Enhanced the image quality of the ROI by over 4 dB.
• Demonstrated significant improvements in applicability and reconstruction quality for real-time GI.

Conclusions:

• The proposed deep learning method significantly boosts GI efficiency and ROI reconstruction quality.
• Enables adaptive ROI selection, paving the way for real-time ghost imaging applications.
• Offers a powerful tool for advancing spectral imaging, 3D imaging, and other GI-based technologies.