Partial hard occluded target reconstruction of Fourier single pixel imaging guided through range slice

Area of Science:

• Optics and photonics
• Computer vision
• Artificial intelligence

Background:

• Fourier single pixel imaging (FSPI) reconstructs images using spatial light modulation and reconstruction algorithms.
• FSPI offers non-locality and high anti-interference but suffers from poor image quality with occluded targets.

Purpose of the Study:

• To enhance Fourier single pixel imaging (FSPI) for reconstructing high-quality images of partially obscured targets.
• To address the limitations of FSPI in scenarios with occlusions and down-sampling.

Main Methods:

• A deep learning-based image inpainting algorithm was developed for FSPI.
• The method incorporates distance-based segmentation for region identification.
• It employs a novel inpainting network (multi-scale sparse convolution and transformer) and a reconstruction network (Channel Attention Mechanism and Attention Gate).

Main Results:

• The proposed method successfully reconstructs complete and clear intensity images from partially obscured targets.
• Demonstrated high inpainting and reconstruction capacity under hard occlusion and down-sampling conditions.
• Significantly improved imaging quality and expanded application scenarios for FSPI.

Conclusions:

• The integrated deep learning approach effectively overcomes occlusion challenges in Fourier single pixel imaging.
• This advancement broadens the utility of FSPI in complex imaging environments.
• The study validates the method's performance through simulations and real-world experiments.