Structural-prior guided bi-generative network for image inpainting

Abstract

Image inpainting is a great challenge when reconstructed with realistic textures and required to enhance the consistency of semantic structures in large-scale missing regions. However, popular structural prior guidance methods primarily rely on the reconstruction of structural features. Due to the Markovian property inherent in purely feedforward architectures, noise undergoes persistent accumulation and propagation in early network layers. Without intermediate feedback mechanisms, minor artifacts in shallow layers would be nonlinearly amplified through successive convolution operations and cannot be timely corrected, thereby hindering the extraction of valid structural information. To this end, we presents a bi-generative network (Bi-GNet) guided by specific semantic structures, including an auxiliary network $N_{\text{s}}$ and an inpainting network $N_{\text{inp}}$. Here $N_{\text{s}}$ provides the structural prior information to $N_{\text{inp}}$ for reconstructing the texture details of images. Additionally, we provide the spatial coordinate attention (SCA) and the adaptive feature filtering (AFF) module to ensure structural consistency and texture plausibility in the reconstructed content. Experiments demonstrate that Bi-GNet significantly outperforms other state-of-the-art approaches on three datasets and achieves good inpainting results on the Mogao Grottoes mural dataset.