Face inpainting is an interesting while challenging task in the fields of computer vision and image processing. In this paper, a parsing based approach is worked out to solve facial image inpainting problems. Clearly, an intact facial image may contain extensive details, and hence a straight forward overall reconstruction is hardly achieved when serious damages exist. To deal with this problem, we choose to first recover the overall image structure represented by parsing images which are quite simple compared with detailed original images. Base on this idea, a two stages based face inpainting framework is proposed, where the first stage exclusively conductss parsing image inpainting, while the following second stage recover all details. Furthermore, a Semantic Compensation Module (SCM) is fused into the first stage to ensure effective context information aggregation, and a Contextual Attention Module (CAM) is brought into the second stage to further improve the appearance rationality. Extensive experiments are conducted on the publicly available CelebA-HQ dataset to verify the effectiveness of the proposed approach.