MFSR: Multi-fractal Feature for Super-resolution Reconstruction with Fine Details Recovery

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Fractal features are crucial for capturing intricate micro and macro texture details in images.
• Existing super-resolution methods often struggle with fine detail recovery, especially in complex textures.
• Diffusion models have shown promise in image generation and restoration tasks.

Purpose of the Study:

• To propose MFSR, a diffusion model-based super-resolution method incorporating multi-fractal features as texture priors.
• To enhance the recovery of fine details in low-resolution images.
• To improve the performance of super-resolution models through effective texture representation.

Main Methods:

• MFSR utilizes a Multi-Fractal Feature Extraction Block (MFB) with Density Estimation (DEB), Similar Feature Grouping (SFGB), Grouped Processing (GPB), and Feature Aggregation (FAB).
• A modified U-Net denoiser integrates these multi-fractal features as reinforcement conditions.
• An attention-based sub-denoiser with Fast Fourier Transform (FFT) is employed to mitigate high-frequency noise during upsampling.

Main Results:

• MFSR achieved superior performance on FFHQ, DIV2K, and Urban100 datasets at 4x resolution.
• On FFHQ, MFSR reached PSNR of 26.94 dB and SSIM of 0.833, outperforming ResDiff.
• Ablation studies confirmed that multi-fractal features consistently improved existing models like SR3, SRDiff, and ResDiff.

Conclusions:

• The proposed MFSR method effectively leverages multi-fractal features for enhanced super-resolution.
• The integration of multi-fractal texture priors significantly improves fine detail recovery.
• MFSR demonstrates generalizability and superior performance compared to existing state-of-the-art methods.