Joint Structure and Texture-guided Image Inpainting with a Multi-Scale Shared Feature Extractor

Significant progress has been made in image inpainting by simultaneously leveraging both structure and texture information for the restoration process. However, there remains considerable room for improvement in effectively fusing and mutually utilizing these structure and texture features. In this paper, we propose a Joint Structure and Texture-guided Image Inpainting model with a Multi-Scale Shared Feature Extractor, which include the Gated Contextual Aggregation (GCA) module, the Adaptive Dual Feature Fusion (ADFF) module, and the Multi-Scale Contextual Attention (MA) module. Specifically, we employ a two-stream network, which models structure-guided texture synthesis and texture-guided structure reconstruction to generate structure and texture information. Concurrently, we introduc structure and texture features. This feature extractor effectively reduces the parameter count while adeptly capturing multi-scale features. Then we utilize the ADFF module to better integrate both types of features, ensuring coherence between structure and texture. Additionally, we employ MA module to refine the fused features, leading to the final inpainting results. Extensive experiments conducted on the CelebA benchmark dataset demonstrate the superiority of the proposed method.