Rethinking Image Inpainting with Attention Feature Fusion

Recent image inpainting models have archived significant progress through learning from large-scale data. However, restoring images under complicated scenarios (e.g. large masks or complex textures) remains challenging. We argue that the inadequate learning of global structure and local texture could lead to the artifacts and blur of current models. Inspired by feature fusion methods, we utilize Attention Feature Fusion (AFF) to better aggregate the different levels of features within our inpainting model from two perspectives. 1) We insert AFF through skip connections to pass long-distance textures to late semantics; 2) Our modified multi-dilated blocks with AFF residual could fuse features in different receptive fields. Both strategies aim to strengthen the texture and structure aggregation and reduce the inconsistency of semantics during learning. We show quantitatively and qualitatively that our approach outperforms current methods on benchmark datasets.