The generative adversarial network (GAN) provide a new way for semantic image inpainting problem. The missing semantic information can be predicted by generating an image with similar distribution of corrupted image based on GAN. In this paper, we propose a high vision quality semantic inpainting algorithm based on a LS-DCGAN. We discuss the optimization of GAN training and introduce the least squares loss function to solve the vanishing gradient problem of DCGAN. Based on a trained LS-DCGAN, we propose a new adversarial loss function for optimizing inpainting network input. Experiment on two datasets show that our algorithm is stable and effective, and have higher naturalness, validity and semantic similarity on visual experience than the state-of-the-art algorithms.