PEPSI++: Fast and Lightweight Network for Image Inpainting

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Generative adversarial network (GAN)-based image inpainting methods achieve high performance.
• Coarse-to-fine networks with contextual attention modules (CAM) are effective but computationally expensive.
• Existing methods suffer from low speed due to numerous parameters and operations.

Purpose of the Study:

• To propose a novel network architecture, PEPSI, for efficient and high-performance image inpainting.
• To reduce hardware costs and improve inpainting quality compared to existing methods.
• To introduce Diet-PEPSI for significant parameter reduction while preserving performance.

Main Methods:

• Developed the parallel extended-decoder path for semantic inpainting (PEPSI) network.
• PEPSI utilizes a shared encoding network and parallel coarse and inpainting decoding paths.
• Introduced rate-adaptive dilated convolutional layers in Diet-PEPSI for efficient global context capture.

Main Results:

• PEPSI and Diet-PEPSI significantly reduce computational time and network parameters.
• Both methods demonstrate improved qualitative scores, including peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).
• Diet-PEPSI achieves substantial parameter reduction with maintained inpainting performance.

Conclusions:

• PEPSI offers an efficient alternative for GAN-based image inpainting, balancing speed and quality.
• Diet-PEPSI provides a highly optimized solution for resource-constrained environments.
• The proposed methods advance the state-of-the-art in semantic image inpainting.