用于离焦图像焦点预测中的深度学习不确定性建模

Uncertainty Modeling in Deep Learning for Focus Prediction of Defocused Images

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±0.46μm和0.83±1.17μm,优于筛选前的0.40±0.66μm和1.08±1.78μm。

tOne of the key technologies in automatic digital microscopy is autofocus.In order to improve the speed of focusing,more and more deep learning methods are being introduced for focus prediction of single-frame images.However,almost all networks believe that their output is necessarily correct,even in the face of unknown samples when the output error results will not include any warning.In this paper,a Bayesian convolutional neural network is proposed to predict the defocus distance from a single image and obtain the uncertainty estimation of the focus prediction results.In addition,uncertainty is proposed to measure the validity of the results,and the focus prediction results are filtered by setting the uncertainty threshold.The proposed method is tested on a large open-source dataset.Experimental results show that the network model proposed in this paper can output higher uncertainty for unknown samples,and the established screening mechanism can effectively reduce the prediction error of the model for unknown samples by eliminating some error results.The model achieved a final error range of 0.37±0.46μm and 0.83±1.17μm on two samples on the public data set,which is better than 0.40±0.66μm and 1.08±1.78μm before screening.