自动数字显微镜的关键技术之一就是自动对焦,为了提升对焦的速度,越来越多的深度学习方法被引入用于单帧图像的焦点预测。然而几乎所有的网络模型都过分信任其输出的结果,面对未知的样本即使输出错误的结果也不会给出任何警示。利用贝叶斯卷积神经网络的实现,可从单张图像中完成离焦距离的预测,并获得焦点预测结果的不确定性估计,此外提出通过设置不确定度阈值实现对焦点预测结果的筛选。在一个大型开源数据集上进行了测试,利用不确定性估计评估预测结果的有效性。结果表明,对比同类型样本,所提出的网络模型在未知样本上能够输出更高的不确定度,建立的筛选机制能有效减小模型在未知样本上的预测误差。在公共数据集上的两个样品的最终误差范围为 0.37±0.46 μm和 0.83±1.17μm,优于筛选前的 0.40±0.66μm和 1.08±1.78μm。
深度学习 自动对焦 贝叶斯神经网络 焦点预测 不确定性分析 deep learning autofocus Bayesian neural network focus prediction uncertainty analysis
Author Affiliations
Abstract
1 Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, Shanghai 201800, China
2 College of Sciences, Jiangnan University, Wuxi 214122, China
The properties of a series of phase measurement techniques, including interferometry, the Hartmann-Shack wavefront sensor, the knife-edge technique, and coherent diffraction imaging, are summarized and their performance in high power laser applications is compared. The advantages, disadvantages, and application ranges of each technique are discussed.
wavefront measurement phase retrieval focus prediction Ptychographical Iterative Engine High Power Laser Science and Engineering
2014, 2(3): 03000e25