基于学习掩模的高速单像素成像

Single pixel imaging (SPI) holds significant promise for addressing specialized imaging challenges, including unconventional wavelength-ranges, proposed scattering scenarios, and low light conditions Recent developments in SPI employing a spinning mask have successfully insured the limits posed by traditional modulators like the Digital Micromirror Device (DMD), specifically with respect to refresh rates and operational spectral bands Never less, current spinning mask implementations, releasing on random patterns or cyclic Hadamard patterns, fall short in achieving rapid and high quality imaging when operating at low sampling rates In this investment, we propose to use deep learning to join optimize a coding scheme based on spinning mask for SPI On the encoding side, a cyclic mask, optimized by the convolutional layer, is meticulously crafted to modify the input object On the coding side, the object image is reconstructed from the modulated intensity fluctuations employing a lightweight neural network integrated with physical model By adapting this approach, we realized residual image results compiling 71x73 pixel images with a sampling rate of 4%, all while maintaining module rate of 2.4MHz. Not only, we have achieved image recording speeds exceeding 12KHz The proposed method dramatically improves the imaging effectiveness of SPI, there catalyzing the practical utilization of SPI in domains such as specialized wavelength imaging and high speed imaging

单像素成像（Single Pixel Imaging, SPI）在应对特定成像难题方面具备显著应用潜力，涵盖非常规波段成像、复杂散射场景以及低光照条件下的成像任务。近期基于旋转掩膜的单像素成像技术进展，成功突破了数字微镜器件（Digital Micromirror Device, DMD）等传统调制器所带来的性能限制，尤其是在刷新率与工作光谱波段层面。然而，当前主流的旋转掩膜成像方案，无论是采用随机图案还是循环哈达玛图案，在低采样率运行时均难以兼顾快速成像与高质量成像效果。在本研究中，我们提出借助深度学习技术，对基于旋转掩膜的单像素成像编码方案进行联合优化。在编码端，我们通过卷积层优化得到的循环掩膜可被精准设计，以调制输入的目标场景；在解码端，我们采用融合物理模型的轻量级神经网络，从经调制的光强波动信号中重建目标图像。通过该方案，我们在采样率仅为4%的条件下，成功生成了分辨率为71×73像素的重建图像，同时维持2.4MHz的调制速率。不仅如此，我们还实现了超过12kHz的图像记录速度。所提方法显著提升了单像素成像的成像性能，进而推动单像素成像在非常规波段成像、高速成像等领域的实际落地应用。

搜集汇总

背景与挑战

背景概述

该数据集聚焦于基于深度学习优化的旋转掩模单像素成像技术，旨在解决传统调制器在刷新率和光谱波段的限制。通过结合卷积层优化的循环掩模和轻量神经网络物理模型，实现了高速、低采样率下的高质量图像重建，调制速率达2.4MHz，成像速度超过12KHz，适用于特殊波长和高速成像场景。数据集包含17个文件，总计932.84 KB，属于物理学领域，发布于2023年。