Photo Realistic Image Completion via Dense Correspondence.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Traditional image completion methods often rely on sparse correspondence, limiting their ability to handle complex deformations.
• Occlusion is a significant challenge in image completion, requiring robust methods to infer missing information.

Purpose of the Study:

• To propose an advanced image completion algorithm leveraging dense correspondence for improved accuracy.
• To address limitations of sparse correspondence methods by introducing a hierarchical approach.

Main Methods:

• A hierarchical PatchMatch method is employed to progressively estimate dense correspondence, capturing subtle image deformations.
• Nearest Neighbor Field (NNF) interpolation is utilized to create smooth and accurate correspondences over occluded areas.
• A color correction algorithm is applied to ensure seamless integration of completed content with the original image.

Main Results:

• The proposed method successfully estimates dense correspondence, even with significant occlusion.
• NNF interpolation effectively bridges occluded regions, enabling accurate content transfer.
• The final color correction step minimizes visual artifacts, enhancing realism.

Conclusions:

• The developed image completion algorithm, based on dense correspondence and hierarchical PatchMatch, achieves state-of-the-art photorealistic results.
• This method offers a significant improvement over traditional sparse correspondence techniques for image inpainting.