Fine-grained restoration of Mongolian patterns based on a multi-stage deep learning network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Heritage Preservation

Background:

• Traditional Mongolian patterns face degradation due to inheritance and preservation challenges.
• Manual restoration is inefficient, time-consuming, and expensive.
• Existing deep learning image restoration methods are unsuitable for complex patterns like Mongolian motifs.

Purpose of the Study:

• To develop an effective deep learning model for restoring damaged Mongolian patterns.
• To address the limitations of current image restoration techniques for culturally significant, complex patterns.
• To provide an efficient digital solution for preserving Mongolian motifs.

Main Methods:

• A multi-stage deep learning network was proposed for Mongolian pattern restoration.
• Stage 1: Pyramid context encoder for global feature learning and restoration.
• Stage 2: Local restoration network using RIC convolution and MPD down-sampling.
• Stage 3: U-Net with attention mechanism for global refinement.

Main Results:

• The proposed model achieved remarkable results in Mongolian pattern repair.
• Performance was evaluated using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), LPIPS, and L1 Loss.
• The method demonstrated superior performance compared to existing techniques.
• Validation on public datasets confirmed the model's wide applicability.

Conclusions:

• The multi-stage network offers an efficient and effective solution for digital restoration of Mongolian patterns.
• The model successfully handles complex textures and rich colors characteristic of Mongolian motifs.
• This research holds significant application prospects for cultural heritage preservation through digital means.