One 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 ± 0.46 μm and 0.83 ± 1.17 μm on two samples on the public data set，which is better than 0.40 ± 0.66 μm and 1.08 ± 1.78 μm before screening.