A Dual Coupled Feature Pyramid for Traditional Paintings Inpainting with Multi-level Semantic Filtering
Biao Yang, Zihan Chen, Yi Zhang · 2025
Recently, deep learning techniques have been widely applied to image generation and inpainting field. However, most existing methods overlooked the importance of structural information, leading to suboptimal results when facing large damaged areas. To compensate for the shortcomings, we develop a dual coupled feature pyramid network in this paper to model the Structure and Texture features (ST features) in a disentangled way. Then, a Feature Integration Module (FIM) is employed to fuse ST features and to generate the inpainted results. Finally, a Multi-Level Semantic Filter (MLSF) is applied to enhance the characterization of semantic details. Experiments have been carried out on 2 datasets, where our network exhibits superior performances than other competitors.