Error-Mask-Adaptive Dynamic Filtering for Image Inpainting

作     者：Ko, Keunsoo Woo, Seunggyun Kim, Chang-Su 

作者机构：Catholic Univ Korea Dept Artificial Intelligence Bucheon 14662 South Korea Korea Univ Sch Elect Engn Seoul 02841 South Korea 

出 版 物：《IEEE ACCESS》 (IEEE Access)

年 卷 期：2025年第13卷

页      面：18403-18417页

核心收录：

基　　金：Catholic University of Korea Samsung Electronics Company Ltd. [IO201214-08156-01] National Research Foundation of Korea (NRF) - Korean Government (MSIT) [RS-2022-NR068986, RS-2024-00397293] 

主　　题：Filtering Convolution Image reconstruction Heuristic algorithms Filtering algorithms Adaptive filters Adaptation models Prediction algorithms Image restoration Visualization Image inpainting dynamic filtering mask-adaptive filtering convolutional neural networks 

摘      要：A novel error-mask-adaptive dynamic filtering (EMDF) algorithm is proposed in this paper, which uses a continuous error mask to inpaint an image adaptively and faithfully. In an EMDF layer, we determine a spatially varying filter adaptively according to an error mask and perform separable dynamic filtering. Meanwhile, we update the error mask by modeling the error propagation during the filtering. Through several EMDF layers, we predict an inpainting result. Finally, we refine it to reconstruct a more faithful image. Experimental results on diverse datasets show that the proposed EMDF algorithm outperforms existing inpainting algorithms significantly. The source codes are available at https://***/keunsoo-ko/EMDF.