Image Inpainting using Clustered Planar Structure Guidance
Emiko Horikawa, Irawati Nurmala Sari, Weiwei Du · 2021
This paper presents an effective method using clustered planar structure guidance for image inpainting. Our method concerns restoring the unknown area by clustering structures of related planes. It is employed to obtain precisely similar structures in the surrounding area of missing regions. The approach of our work contains four essential steps: Planar Guidance, Clustering Structures, Feature Localization, and Patch Matching. According to perspective scenes, we first extract vanishing points (vp1, vp2, and vp3) using RAndom SAmple Consensus (RANSAC) algorithm as planar guidance. Then, we cluster the structure lines of each planar into more categories using the integration between K-means++ and Elbow method. Gaussian filter and Hadamard products blend among structure categories in the feature localization. This feature position propagates the surrounding structure information into unknown areas. For completing the unknown area, we employ PatchMatch [2] algorithm to match between the unknown and its surrounding patches. In experiments, our results perform well in various structures and perspective scenes that have foreshortened planes.