Guiding image inpainting via structure and texture features with dual encoder

Image inpainting techniques have made rapid progresses in recent years. Recent advancements focus mainly on generating realistic and semantically plausible structure and texture features in missing regions. However, current popular inpainting methods rely typically on a single encoder–decoder or two separate encoder–decoders, which lead to inconsistent contextual semantics and blurry textures. To address the above issue, a dual-feature encoder implemented by structure and texture features is proposed. It utilizes skip connection to guide its corresponding decoder to fill image structure information (deep layer) and texture information (shallow layer) in the attention-based latent space. Additionally, we design multi-scale receptive fields to further improve the consistency of contextual semantics and image details. The experimental findings demonstrate that our method can effectively repair the structure and texture information of missing images with superior performance on three commonly used datasets. Furthermore, we build a mural dataset from the Mogao Grottoes and successfully restore them using our network.