PEIPNet: Parametric Efficient Image-Inpainting Network with Depthwise and Pointwise Convolution

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Current image inpainting research often increases model complexity and resource demands.
• This trend poses challenges for computational efficiency and practical application.

Purpose of the Study:

• To propose a Parametric Efficient Image Inpainting Network (PEIPNet) for efficient and effective image inpainting.
• To address the computational burden associated with traditional inpainting methods.

Main Methods:

• Developed a one-stage inpainting framework utilizing depthwise and pointwise convolutions to minimize parameters and computational cost.
• Integrated spatially-adaptive denormalization (SPADE) for mask-conditional normalization, dense dilated convolution module (DDCM) to prevent gradient vanishing and capture global context, and efficient self-attention (ESA) for long-range information extraction.

Main Results:

• PEIPNet demonstrates significantly lower operational memory usage compared to existing state-of-the-art methods.
• Qualitative and quantitative experiments confirm the model's generalized inpainting capabilities across diverse datasets (Paris StreetView, CelebA, Places2).

Conclusions:

• PEIPNet offers an efficient and effective solution for image inpainting tasks.
• The proposed architecture successfully balances performance with reduced computational resource requirements.