In this paper, an unsupervised deep learning algorithm is presented to solve the problem of multi-focus fusion of rock slice images collected under a microscope. In order to extract the deep features of images, a codec neural network is trained with unsupervised method to extract the depth features of different focused images and get the feature map. Then, a binary decision graph is calculated using the spatial frequency of the feature graph. Due to subtle decision bias, there may be holes and burrs in the binary decision maps, so the decision maps are morphologically processed and filtered. Finally, the fused image is obtained from the processed decision map. The experimental results show that the data evaluation index 、、 of this method is 0.7477、0.9874、0.7969.At the same time, the subjective effect is better than other methods. Therefore, Experiments show that this method can achieve good results in the application of multi-focus fusion of microscopic rock slice images and general images.