Position and Channel Attention for Image Inpainting by Semantic Structure
Jingjun Qiu, Yan Gao · 2020
Image inpainting has made great progress with the emergence of deep learning. However, there are also problems of structural loss and blurry textures, which lead to the generated image has artifacts and incomplete object. When a heavily structured object is partially masked, most methods cannot complete the repair work while keeping the semantic structure intact. To solve these problems, we propose a two-stage adversarial model: introduce unsupervised semantic structure guidance. Compared with other methods, the semantic information obtained by the introduction of semantic segmentation is more accurate. Compared with the structure information of supervised semantic segmentation of manual labels, the unsupervised semantic structure information is more flexible, and the labels are more diverse. The attention model with location information and channel information strengthens the model's long-range contextual information and multi-scale context information fusion capabilities. We evaluate our model over the publicly available datasets CelebA, Places2, and Paris StreetView, our method has a higher repair quality than the existing state-of-the-art approaches, especially when repairing the image with the large missing area.