Image Inpainting With Local and Global Refinement.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning has advanced image inpainting, typically using encoder-decoder networks with large receptive fields.
• However, large receptive fields are not always optimal for image inpainting, especially for local structures and textures, and can introduce undesired completion artifacts.

Purpose of the Study:

• To propose a novel three-stage image inpainting framework that rethinks the receptive field's role.
• To introduce a method that balances local and global refinement for improved inpainting performance.

Main Methods:

• A three-stage framework combining coarse initial inpainting, local refinement with a small receptive field, and global refinement with an attention-based large receptive field.
• Utilizing an encoder-decoder network with skip connections for initial results, followed by shallow and attention-based deep models for refinement.

Main Results:

• The proposed method outperforms state-of-the-art approaches on three public image inpainting datasets.
• The local and global refinement network can be integrated into existing models to enhance their performance.

Conclusions:

• The proposed three-stage framework effectively addresses limitations of fixed large receptive fields in image inpainting.
• The novel approach achieves superior results by incorporating both local and global refinement strategies, offering a versatile enhancement for existing inpainting networks.