RNON: image inpainting via repair network and optimization network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models offer advantages in image inpainting over traditional methods, generating better structure and texture.
• Existing convolutional neural network methods often suffer from color differences, texture loss, and distortion.

Purpose of the Study:

• To propose an effective image inpainting method addressing limitations of current deep learning techniques.
• To improve visual effects and image quality in repaired images.

Main Methods:

• A novel image inpainting method utilizing two independent generative adversarial networks (GANs).
• An image repair network using a partial convolutional network for irregular missing areas.
• An image optimization network employing deep residual networks to correct chromatic aberration.

Main Results:

• The proposed method, RNON, demonstrates superior performance in qualitative and quantitative evaluations.
• Synergistic operation of the two network modules significantly improves visual effects and image quality.
• RNON effectively repairs irregular missing image areas and corrects local chromatic aberrations.

Conclusions:

• The proposed RNON method offers an effective solution for image inpainting, outperforming state-of-the-art approaches.
• The dual-network architecture successfully addresses color differences and texture distortion issues.
• This research advances the field of image inpainting through improved generative adversarial network application.