In recent years, more and more researchers use deep learning to process inpainting tasks. Among them, the use of generation countermeasure network to process inpainting tasks has become more and more popular and has achieved good results. However, there are still issues with blurry repair results and unsmooth structure. In this paper, we propose a method of inpainting based on u-net structure for generation adversarial network, the first two layers of our encoder use multi-scale shallow feature extraction modules (MSFEM) to extract lowdimensional texture and structural information. We introduce multi-scale spatial attention module (MSAM) into skip connections to obtain more shallow features and improve repair performance. The decoder uses improved dense convolutional blocks to fully utilize and extract feature information. The experiment used two datasets, CelebA and Palace2, through experiments, the repair effect of our proposed method is better than the state-of-the-art image inpainting approaches.