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

Kangning Zhang, Yifei Sun, Varun Yelluru, Weijian Yang

Abstract: Single-pixel imaging (SPI) within the framework of compressive sensing (CS) is a powerful technique that enables image acquisition at sub-Nyquist sampling rates by leveraging the sparse latent representations of the object scenes. As a cost-effective alternative to focal plane array cameras, SPI has been explored for various imaging applications. We recently introduced a novel block-scanning SPI approach that samples the scene using a single, learnable illumination pattern, which substantially enhanced acquisition speed compared to traditional SPI systems that rely on pattern switching via digital micromirror devices. In this work, we present a new computational framework, termed Incoherence-Enhanced Generative Adversarial Deep Compressive Imaging (IE-GADCI), designed to jointly optimize both the illumination pattern and the image reconstruction algorithm for block-scanning SPI, under the principle of compressive sensing. Our architecture employs a neural network that learns the latent sparse representations of the scene and integrates information from both the image and sparsity domains to achieve high-resolution reconstructions with high computational efficiency. A key innovation of IE-GADCI is its optimization of the incoherence between the illumination pattern and the sparse representation, which substantially improves reconstruction fidelity. We validated the performance of IE-GADCI through numerical simulations on natural and biomedical image datasets. We benchmarked IE-GADCI against state-of-the-art methods used in SPI, and additionally, single-image super-resolution (SISR), given the conceptual similarity between block-scanning SPI and SISR. At a subsampling rate of just 1.5625%, IE-GADCI achieves a peak signal-to-noise ratio (PSNR) exceeding that of competing methods by more than 2 dB. These results highlight the potential of IE-GADCI for high-speed, high-fidelity imaging in applications such as consumer electronics and biomedical imaging, including calcium imaging for neuronal activity recording.