Block-based compressive imaging with a swin transformer

Area of Science:

• Optics and photonics
• Computer vision
• Machine learning

Background:

• Block-based compressive imaging (BCI) enables high-speed sampling using a spatial light modulator and low-resolution detector.
• BCI reduces computational load compared to traditional compressive imaging but can introduce block artifacts.
• Super-resolution algorithms are crucial for reconstructing high-quality images from BCI data.

Purpose of the Study:

• To develop an advanced deep neural network for improved block-based compressive imaging reconstruction.
• To address and mitigate block artifacts inherent in BCI.
• To achieve real-time, high-quality image reconstruction in BCI systems.

Main Methods:

• Proposed SwinBCI, a data-driven deep neural network leveraging the swin transformer architecture.
• Incorporated local attention and shifted window mechanisms for enhanced reconstruction.
• Utilized a dataset for model training to acquire prior knowledge.
• Employed graphics processing unit (GPU) acceleration for reduced computation time.
• Investigated cake cutting-Hadamard matrix sampling for improved performance.

Main Results:

• SwinBCI demonstrated superior image reconstruction quality compared to traditional methods.
• The model achieved significantly reduced computation times, enabling real-time BCI.
• Cake cutting-Hadamard matrix sampling yielded better reconstruction results than Bernoulli matrix sampling.
• Experimental validation on diverse datasets and actual BCI systems confirmed SwinBCI's effectiveness.

Conclusions:

• SwinBCI offers a powerful solution for high-quality and fast image reconstruction in block-based compressive imaging.
• The integration of swin transformers and GPU acceleration facilitates real-time BCI applications.
• The proposed method outperforms existing compressed sensing reconstruction techniques across various sampling rates.