Thanka Mural Inpainting Based on Multi-Scale Adaptive Partial Convolution and Stroke-Like Mask.

Area of Science:

• Art Conservation
• Computer Vision
• Digital Heritage

Background:

• Thanka murals are vital Tibetan cultural heritage, but historical damage threatens their preservation.
• Existing partial convolution methods struggle with multi-scale features, stroke-like patterns, and accurate content restoration in Thanka murals.

Purpose of the Study:

• To develop an advanced inpainting method for restoring damaged Thanka murals.
• To address limitations of current partial convolution techniques in mural restoration.

Main Methods:

• Proposed a multi-scale adaptive partial convolution (MAPConv) for accurate feature extraction.
• Introduced parameter-configurable stroke-like masks to simulate Thanka inpainting patterns.
• Implemented a 2-phase learning framework using MAPConv U-Net and specialized loss functions.

Main Results:

• The method successfully restored damaged Thanka murals on a small dataset (N=2780).
• Achieved high-speed restoration (600 ms for multiple holes in 512x512 images).
• Generated realistic mural content, demonstrating effectiveness for simulated and real damages.

Conclusions:

• The proposed end-to-end method effectively restores Thanka murals by overcoming limitations of existing techniques.
• The approach shows promise for application in other small dataset inpainting tasks.
• This work contributes to the digital preservation of cultural heritage through advanced AI.