Abstract: Existing image inpainting methods based on deep learning have the problem of insufficient utilization of multi-scale feature space information in undamaged areas of images. For this problem, we proposed an image inpainting model based on pyramid spatial attention and feature reasoning. Firstly, the region identification module based on partial convolution is used to identify the region that needs reasoning in this cycle. Secondly, the circular feature inference module is used to efficiently reason the image features of the area to be reasoned. Finally, the feature fusion module based on residual redundancy feature is used to ensure that the interference of invalid feature information on image inpainting is reduced in the process of fusing intermediate feature maps. We experiment with our model end-to-end on the face, street view, and other datasets, and the results show that compared with the classical methods, the peak signal-to-noise ratio,structural similarity, and average L₁ loss evaluation index of this model are improved by 3%, 1%, and 3%,respectively.