A dual branch attention network based on practical degradation model for face super resolution

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Face super-resolution (FSR) struggles with real-world image degradation.
• Existing FSR methods lack generalization due to inadequate degradation modeling.
• Complex degradations like noise and blur severely impact FSR performance.

Purpose of the Study:

• To propose a practical degradation model for realistic face image distortion simulation.
• To develop a robust face super-resolution network capable of handling diverse degradations.
• To enhance the generalization ability and reconstruction performance of FSR methods.

Main Methods:

• A novel hybrid degradation model employing a stochastic strategy for multiple degradation processes (Gaussian noise, Rayleigh noise, Motion blur, Salt-and-Pepper noise, Mean blur).
• Design of a dual-branch attention network (DBANet) for face super-resolution.
• Experimental validation on SCUT_FBP, Helen, and PFHQ datasets.

Main Results:

• DBANet demonstrates effectiveness in handling various image distortion modalities.
• The proposed model achieves satisfactory results on benchmark datasets.
• Improved robustness in super-resolution reconstruction under complex degradations.

Conclusions:

• The hybrid degradation model and DBANet offer a significant advancement in face super-resolution.
• The approach enhances FSR robustness and generalization for real-world applications.
• This work provides a foundation for future research in robust image super-resolution.