Local Texture Pattern Estimation for Image Detail Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Deep super-resolution (SR) networks struggle with restoring random high-frequency textures.
• Generative adversarial networks (GANs) excel at texture restoration but have drawbacks like large parameters and potential for fake textures.

Purpose of the Study:

• To propose a novel SR network for restoring fine high-frequency texture details without relying on GANs.
• To address the limitations of existing SR methods in accurately reconstructing complex textures.

Main Methods:

• Developed a novel SR network based on local texture pattern estimation (LTPE).
• Designed a differentiable local texture operator to extract texture structures.
• Implemented a texture enhancement branch for predicting high-resolution local texture distribution.
• Utilized a texture fusion SR branch guided by the estimated texture map.
• Optimized the network using L1 loss and Gram loss.

Main Results:

• The proposed LTPE-based SR network effectively recovers high-frequency textures.
• The method achieves high-quality texture reconstruction without GAN structures.
• Restored high-frequency details are constrained by local texture distribution, reducing generation errors.

Conclusions:

• The LTPE-based SR network offers a viable alternative to GANs for realistic texture restoration.
• This approach improves the accuracy and fidelity of high-frequency texture recovery in image SR.