【摘要】 针对稀疏投影下保偏光纤轴向残余应力重建时存在的伪影噪声问题，提出一种基于深度学习的正则化迭代重建算法。基于光弹原理与断层扫描分析光纤轴向残余应力的二维分布重建，在稀疏投影测量中，现有的滤波反投影法实现Radon逆变换进行应力重建存在伪影噪声问题；分析保偏光纤轴向残余应力测量的离散投影模型，针对稀疏投影导致的不适定问题，引入广义正则项，分析了应力的正则化迭代重建；通过深度学习，从样本中学习广义正则项的具体形式，并用训练好的模型进行仿真实验。对5种保偏光纤的重建结果表明：该方法在稀疏投影测量中同时具备很好的伪影噪声抑制与细节特征保持能力，在投影数目为36个视角时均方根误差为0.198 5 MPa、峰值信噪比为47.344 5 dB、结构相似性指数达到0.995 1,分别提升59.7%、18.4%、1.34%;单个样本的重建时间低于1 s,可以实现应力的快速、准确重建。更多还原

【Abstract】 To address the issue of artifact noise during the reconstruction of axial residual stress in polarization-maintaining fiber under sparse-view, a deep learning-based regularized iterative reconstruction algorithm was proposed.Based on the principles of photoelasticity and tomography, the two-dimensional distribution of axial residual stress in fiber reconstruction was analyzed.In sparse-view measurement, the existing filtered back projection method that implemented Radon inversion for stress reconstruction suffered from artifact noise issues.The discrete projection model of the measurement of axial residual stress in polarization-maintaining fiber was analyzed, and considering the ill-posed problem caused by sparse-view sampling, a generalized regularization term was introduced, and the regularization iterative reconstruction of stress was analyzed.The specific form of the generalized regularization term was learned from samples through deep learning, and simulation experiments were conducted using the trained model.The reconstruction results for five types of polarization-maintaining fibers show that this method possesses excellent capabilities in suppressing artifact noise and preserving detail features in sparse-view measurements.With 36 projection angles, the root mean square error is 0.1985 MPa, the peak signal-to-noise ratio is 47.344 5 dB,and the structural similarity index reaches 0.995 1,representing improvements of 59.7%,18.4%,and 1.34% respectively.The reconstruction time for a single sample is less than 1 second, allowing for rapid and accurate reconstruction of stress.更多还原