Pseudo Decoder Guided Light-Weight Architecture for Image Inpainting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image inpainting synthesizes missing image regions for applications like object removal and virtual try-on.
• Existing methods often rely on computationally expensive coarse-to-fine architectures or external guided information (edges, structures).

Purpose of the Study:

• To propose a computationally efficient and lightweight network for image inpainting.
• To develop an inpainting method that does not require guided information, reducing complexity and resource demands.

Main Methods:

• A novel architecture featuring a multi-encoder level feature fusion module, a pseudo decoder, and a regeneration decoder.
• The multi-encoder fusion merges structural and textural information from diverse receptive fields.
• A space-depth correlation module aids the regeneration decoder for accurate inpainting.

Main Results:

• The proposed network achieves effective image inpainting with a minimal parameter count (0.97M).
• Demonstrated state-of-the-art performance on benchmark datasets (Paris Street View, Places2, CelebA_HQ) across various mask types.
• Successfully tested on high-resolution images (1024x1024, 2048x2048).

Conclusions:

• The proposed lightweight network offers a computationally efficient solution for image inpainting.
• Effectiveness is validated through extensive comparisons, complexity analysis, and ablation studies.
• Presents a viable alternative to resource-intensive inpainting methods.