FreNTS: Neural Texture Synthesis in Frequency Domain

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing texture synthesis methods struggle with regular, densely interconnected structures.
• Neural texture synthesis often faces challenges in maintaining structural integrity and realism.

Purpose of the Study:

• To propose a novel neural texture synthesis method, FreNTS, for generating realistic regular textures.
• To enhance texture synthesis by incorporating frequency domain information and adaptive guided correspondence loss.

Main Methods:

• Utilizing Discrete Cosine Transform (DCT) on image patches to extract frequency domain features.
• Implementing an adaptive guided correspondence (AGC) loss function for joint spatial and frequency domain optimization.
• Introducing Tile LPIPS as a quantitative evaluation metric for texture synthesis quality.

Main Results:

• FreNTS effectively synthesizes textures with continuous, complete, and visually realistic structures.
• The method accelerates neural texture synthesis and captures superior structural details using high-frequency information.
• Experimental results validate the efficacy of FreNTS in generating high-quality regular textures.

Conclusions:

• FreNTS offers a significant advancement in neural texture synthesis, particularly for regular textures.
• The integration of frequency domain analysis and AGC loss leads to improved realism and efficiency.
• The proposed method addresses limitations of previous approaches in synthesizing complex structural patterns.