Learning-based adaptive under-sampling for Fourier single-pixel imaging.

Area of Science:

• Computational Imaging
• Optical Physics
• Machine Learning

Background:

• Fourier single-pixel imaging (FSI) traditionally requires extensive data acquisition.
• Efficient reconstruction from undersampled measurements is crucial for practical FSI applications.
• Optimizing both sampling strategy and reconstruction algorithm simultaneously presents a significant challenge.

Purpose of the Study:

• To develop an efficient, learning-based method for Fourier single-pixel imaging (FSI).
• To introduce an adaptive undersampling technique (AuSamNet) that optimizes sampling masks and deep neural networks concurrently.
• To enable high-quality image reconstruction from highly undersampled Fourier spectrum data.

Main Methods:

• Developed AuSamNet, an auto-encoder-based framework for adaptive undersampling in FSI.
• Co-optimized a deep neural network and sampling mask for efficient data acquisition and reconstruction.
• Validated the method through simulations and experimental reconstructions of natural color images.

Main Results:

• AuSamNet successfully reconstructed high-quality natural color images at a sampling ratio as low as 7.5%.
• The method demonstrated superior performance compared to traditional FSI techniques under severe undersampling.
• The adaptive undersampling strategy proved effective for optimizing the FSI encoding and decoding scheme.

Conclusions:

• The proposed AuSamNet method significantly enhances the efficiency of Fourier single-pixel imaging.
• This adaptive undersampling strategy holds potential for application in other computational imaging modalities like tomography and ptychography.
• The developed source code is publicly available to facilitate further research and development.