TomoSphero：用于球面网格上行星与太阳层析成像的快速可微投影器
TomoSphero: Fast Differentiable Projector for Planetary and Solar Tomography on Spherical Grids
浏览论文内容
中文总结 AI 辅助
本文提出基于PyTorch实现的TomoSphero，一种适用于行星与太阳层析成像球面网格的快速可微投影器，含常见投影类型且灵活适配任意投影，可用于重建算法并支持GPU加速与自动微分。
中文摘要 AI 辅助
计算层析成像是一种通过一组投影（通常沿某一规则路径获取）确定物体内部结构的工具。近年来，已出现可从沿更复杂路径的投影快速重建的方法及GPU加速库，但这些库大多依赖物体的笛卡尔离散化，并非适用于所有场景。本文提出TomoSphero，这是一种用于球面网格的可微层析投影器，球面网格常用于行星与太阳层析成像。TomoSphero设计为重建算法的构建模块，包含锥束、平行束等常见投影类型，且足够灵活以适配任意投影。TomoSphero基于PyTorch实现，支持在GPU上快速计算投影，可便捷使用现代机器学习优化器，还具备自动微分功能，用于参数化模型的快速原型开发。
英文摘要
Computational tomography is a tool for determining the internal structure of objects from a set of projections, typically taken along some regular path. In recent years, methods and GPU-accelerated libraries have emerged that allow for fast reconstruction from projections along more complicated paths. Most of these libraries rely on a Cartesian discretization of the object, which is not appropriate for all scenarios. We present TomoSphero, a differentiable tomographic projector over spherical grids which are often used in planetary and solar tomography. TomoSphero is designed to be used as a building block in reconstruction algorithms and includes common projection types such as cone-beam and parallel-beam, but is flexible enough to accommodate arbitrary projections. TomoSphero is implemented in PyTorch which allows for fast projection computation on GPUs, easy access to modern machine learning optimizers, and automatic differentiation for rapid prototyping of parametric models.
