Noise Robust Face Image Super-Resolution Through Smooth Sparse Representation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Face image super-resolution is a significant area of research.
• Sparse Representation (SR)-based methods offer competitive performance for super-resolution.
• Existing SR methods struggle with high noise levels, leading to unstable reconstructions.

Purpose of the Study:

• To develop a robust face image super-resolution method that performs well under high noise conditions.
• To enhance the stability of reconstruction weights in SR-based approaches when dealing with noisy low-resolution (LR) images.

Main Methods:

• Proposed a novel SR-based face super-resolution approach incorporating smooth priors.
• Introduced fused least absolute shrinkage and selection operator (LASSO)-based smooth constraint.
• Implemented locality-based smooth constraint within a least squares representation framework.

Main Results:

• The proposed method achieved stable reconstruction weights even with high noise levels.
• Experimental results on FEI and CMU+MIT face databases demonstrated superior performance.
• Visual and quantitative comparisons confirmed enhanced reconstruction quality for noisy LR face images.

Conclusions:

• The novel SR-based approach with smooth priors effectively addresses the limitations of existing methods in noisy environments.
• The incorporation of LASSO and locality constraints leads to more stable and accurate face image super-resolution.
• This method offers a significant improvement for reconstructing high-resolution faces from heavily corrupted low-resolution inputs.