Structure-Texture Consistent Painting Completion for Artworks
Irawati Nurmala Sari, Weiwei Du
Abstract
Image completion techniques have made rapid and impressive progress due to advancements in deep learning and traditional patch-based approaches. The surrounding regions of a hole played a crucial role in repairing missing areas during the restoration process. However, large holes could result in suboptimal restoration outcomes due to complex textures causing significant changes in color gradations. As a result, they led to errors such as color discrepancies, blurriness, artifacts, and unnatural colors. Additionally, recent image completion approaches focused mainly on scenery and face images with fewer textures. Given these observations, we present a structure-texture consistent completion approach for filling large holes with detailed textures. Our method focuses on improving image completion in the context of artworks, which are expressions of creativity and often have more diverse structures and textures from applying paint to a surface using brush strokes. To handle the unique challenges posed by artwork, we segment non-homogeneous areas and then use Cohesive Laplacian Fusion to complete the texture of the remaining missing segmented area. This technique involves detecting changes in base structures and textures using multiple matched patches to achieve more consistent results. The experimental results show that our proposed method is competitive and outperforms state-of-the-art methods in missing regions and color gradations of art paintings.