IE-GADCI: An End-to-End Incoherence-Enhanced Generative Adversarial Deep Compressive Imaging.

Area of Science:

• Computational imaging
• Compressive sensing
• Machine learning for imaging

Background:

• Single-pixel imaging (SPI) offers a cost-effective alternative to focal plane array cameras for image acquisition.
• Traditional SPI relies on pattern switching, limiting acquisition speed.
• Block-scanning SPI with learnable illumination patterns enhances speed but requires optimized reconstruction.

Purpose of the Study:

• To develop a novel computational framework, IE-GADCI, for joint optimization of illumination patterns and reconstruction algorithms in block-scanning SPI.
• To improve reconstruction fidelity and computational efficiency in single-pixel imaging.
• To enhance the performance of compressive sensing (CS) based imaging systems.

Main Methods:

• Developed Incoherence-Enhanced Generative Adversarial Deep Compressive Imaging (IE-GADCI) framework.
• Employed a neural network to learn scene sparse representations and integrate image/sparsity domain information.
• Optimized the incoherence between illumination patterns and sparse representations.

Main Results:

• IE-GADCI achieved high-resolution reconstructions with high computational efficiency.
• Demonstrated significant improvement in reconstruction fidelity by optimizing pattern-sparsity incoherence.
• At 1.5625% subsampling, IE-GADCI surpassed competing methods by over 2 dB PSNR.

Conclusions:

• IE-GADCI offers a powerful approach for high-speed, high-fidelity block-scanning SPI.
• The framework shows potential for applications in consumer electronics and biomedical imaging, including calcium imaging.
• Joint optimization of illumination and reconstruction is crucial for advanced compressive imaging.