Noise Robust Face Hallucination Based on Smooth Correntropy Representation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Sparse manifold learning (SML) methods are popular for face hallucination but struggle with noisy images due to least-square regression (LSR).
• Existing SML approaches lack robustness against noise, limiting their effectiveness in real-world scenarios.

Purpose of the Study:

• To propose a novel Smooth Correntropy Representation (SCR) model for robust noisy face hallucination.
• To enhance the resolution of low-resolution (LR) face images corrupted by noise.

Main Methods:

• Developed a unified framework combining correntropy regularization and smooth constraints for face hallucination.
• Introduced the correntropy induced metric (CIM) to replace LSR for noise-robust error approximation.
• Incorporated a fused LASSO penalty in the feature space to preserve the manifold structure of similar training samples.

Main Results:

• The proposed SCR model demonstrates superior performance in super-resolving noisy LR face images compared to state-of-the-art methods.
• SCR effectively handles noise with uncertain distributions, maintaining accuracy in degraded image conditions.
• The method successfully exploits the inherent typological structure of patch manifolds for more accurate representations.

Conclusions:

• The Smooth Correntropy Representation (SCR) model offers a robust and effective solution for noisy face hallucination.
• SCR significantly improves image resolution and representation accuracy in the presence of noise.
• This approach advances the field of image super-resolution for degraded facial imagery.