基于残差编解码网络的单光子压缩成像

Single-Photon Compressive Imaging Based on Residual Codec Network

在进行高分辨率成像时,由于需要大量的测量和图像重建计算,单光子压缩成像需要较长的时间。提出了一种采样和重建集成的残差编解码网络SRIED-Net用于单光子压缩成像。将二值化的全连接层作为网络的第一层,并将其训练成二进制的测量矩阵,直接加载到数字微镜阵列上以实现高效压缩采样。除第一层外的其余网络都用于快速重建压缩感知图像。通过一系列的仿真和系统实验比较了压缩采样率、测量矩阵和重建算法对成像性能的影响。实验结果表明,SRIED-Net在低测量率下优于目前比较先进的迭代算法TVAL3,在高测量率下与TVAL3的效果很接近,在所有测量率下都优于目前常见的几种基于深度学习的方法。

When performing high-resolution imaging using a single-photon compressive technique,a long imaging time is required owing to numerous measurements and a large number of image-reconstruction calculations.We demonstrate a sampling-and-reconstruction-integrated residual codec network,namely SRIED-Net,for single-photon compressive imaging.We use the binarized fully connected layer as the first layer of the network and train it into a binary-measurement matrix to directly load onto the digital micromirror device for efficient compressive sampling.The remaining layers of the network are used to quickly reconstruct the compressed sensing image.We compare the effects of the compressive sampling rate,measurement matrix,and reconstruction algorithm on imaging performance through a series of simulations and system experiments.The experimental results show that SRIED-Net is superior to the current advanced iterative algorithm TVAL3 at a low measurement rate and that its imaging quality is similar to that of TVAL3 at a high measurement rate.It is superior to current deep-learning-based methods at all measurement rates.