A Broad Generative Network for Two-Stage Image Outpainting.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image outpainting, generating large images from patches, is computationally intensive.
• Existing two-stage frameworks offer step-by-step solutions but suffer from long training times.
• Optimizing network parameters in deep learning is challenging with limited iterations.

Purpose of the Study:

• To propose a novel Broad Generative Network (BG-Net) for efficient two-stage image outpainting.
• To accelerate the training process of two-stage image outpainting frameworks.
• To enhance the quality and reduce the training duration of image outpainting methods.

Main Methods:

• Developed a Broad Generative Network (BG-Net) as a reconstruction network for the first stage.
• Utilized ridge regression optimization for rapid training of the reconstruction network.
• Introduced a seam line discriminator (SLD) in the second stage for transition smoothing.

Main Results:

• BG-Net achieved state-of-the-art results on Wiki-Art and Place365 datasets, outperforming existing methods.
• The method demonstrated superior performance in Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) metrics.
• Achieved faster training speeds compared to deep learning-based networks, reducing overall training duration to one-stage levels.

Conclusions:

• The proposed BG-Net offers a powerful and efficient solution for image outpainting.
• The method significantly improves image quality through effective transition smoothing.
• BG-Net demonstrates adaptability to recurrent outpainting, showcasing strong associative drawing capabilities.