High-Resolution Art Painting Completion using Multi-Region Laplacian Fusion

Abstract

Image completion has made impressive advancements based on deep learning approaches. However, even with advanced deep learning such as Generative Adversarial Networks (GAN), the restored area is not always optimal due to small-scale texture synthesis in high resolution and inferring missing information about image content from distant contexts, resulting in distorted lines and unnatural colors, especially in art painting completion with complicated structures and textures. Although several precious art paintings have been well-preserved by curators in museums, some frequent damages such as scratches, torn-out areas, and holes are still visible and require challenging physical repairs. Therefore, for practical refinement, some researchers convert them into high-resolution digital paintings to generate crisp brush strokes, textures, shapes, and tones by assuming similarities with the original physical ones. Based on these observations, we propose proceeding with a high-resolution art painting completion by applying a superior traditional method, named Multi-Region Laplacian Fusion. We attempt to recover irregular missing regions expected as the damages of ordinary paintings that often occur. To address high-resolution inpainting, we integrate two completions using the Laplacian pyramid and patch-based propagation. We then apply Alpha blending among both results to yield the fused reaction completion. Our experiments firmly validate the effectiveness of our proposed method to complete art paintings with random irregular missing regions.