Opt Lett. 2022 Nov 1;47(21):5461-5464. doi: 10.1364/OL.472680.
The speed of single-pixel imaging (SPI) is tied to its resolution, which is positively related to the number of modulation times. Therefore, efficient large-scale SPI is a serious challenge that impedes its wide applications. In this work, we report a novel, to the best of our knowledge, sparse SPI scheme and corresponding reconstruction algorithm to image target scenes at above 1 K resolution with reduced measurements. Specifically, we first analyze the statistical importance ranking of Fourier coefficients for natural images. Then the sparse sampling with a polynomially decending probability of the ranking is performed to cover a larger range of the Fourier spectrum than non-sparse sampling. The optimal sampling strategy with suitable sparsity is summarized for the best performance. Next, a lightweight deep distribution optimization (DO) algorithm is introduced for large-scale SPI reconstruction from sparsely sampled measurements instead of a conventional inverse Fourier transform (IFT). The DO algorithm empowers robustly recovering sharp scenes at 1 K resolution within 2 s. A series of experiments demonstrate the technique's superior accuracy and efficiency.
单像素成像 (SPI) 的速度与其分辨率有关，而分辨率又与调制次数成正比。因此，高效的大规模 SPI 是一个严重的挑战，阻碍了其广泛应用。在这项工作中，我们报告了一种新颖的、据我们所知的稀疏 SPI 方案和相应的重建算法，用于以减少的测量值对目标场景进行高于 1 K 分辨率的成像。具体来说，我们首先分析了自然图像傅里叶系数的统计重要性排序。然后，以概率呈多项式递减的方式进行稀疏采样，以覆盖比非稀疏采样更大的傅里叶谱范围。总结了最佳性能的最优稀疏采样策略。接下来，引入了一种轻量级的深度分布优化 (DO) 算法，用于从稀疏采样测量值进行大规模 SPI 重建，而不是传统的逆傅里叶变换 (IFT)。DO 算法能够在 2 秒内稳健地恢复 1 K 分辨率的清晰场景。一系列实验证明了该技术具有优越的准确性和效率。
