Structure-Guided Image Completion with Image-level and Object-level Semantic Discriminators