Interferometry can measure the shape or the material density of a system that could not be measured otherwise by recording the difference between the phase change of a signal and a reference phase. This difference is always between $-\pi$ and $\pi$ while it is the absolute phase that is required to get a true measurement. There is a long history of methods designed to recover accurately this phase from the phase "wrapped" inside $]-\pi,\pi]$. However, noise and under-sampling limit the effectiveness of most techniques and require highly sophisticated algorithms that can process imperfect measurements. Ultimately, analysing successfully an interferogram amounts to pattern recognition, a task where radial basis function neural networks truly excel at. The proposed neural network is designed to unwrap the phase from two-dimensional interferograms, where aliasing, stemming from under-resolved regions, and noise levels are significant. The neural network can be trained in parallel and in three stages, using gradient-based supervised learning. Parallelism allows to handle relatively large data sets, but requires a supplemental step to synchronized the fully unwrapped phase across the different networks.
翻译:干涉测量方法可以通过记录信号和参考阶段的阶段变化之间的差别来测量一个无法以其他方式测量的系统的形状或物质密度。 这种差别总是在$-pi美元和$/pi美元之间,而这是真正测量所需要的绝对阶段。 长期以来,设计了各种方法,以便精确地从$- pi,\pi,\pi,\pi]美元内的“包装”阶段中恢复这个阶段。 但是,噪音和下取样限制了大多数技术的有效性,需要非常复杂的算法来处理不完善的测量。 最终, 分析一个中间图相当于模式的识别, 也就是辐射基功能神经网络真正具有优势的任务。 拟议的神经网络的设计是要将这个阶段从二维的干涉图中解开, 在那里, 由解解后的区域和噪音水平是巨大的。 神经网络可以平行和分三个阶段进行训练, 使用基于梯度的监督下的学习。 平行学可以处理相对大的数据集, 但需要补充步骤, 在不同的网络上同步完全封闭的阶段。