Image interpolation via regularized local linear regression.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Linear regression is effective for image interpolation.
• Ordinary least squares (OLS) optimization in existing methods lacks robustness to outliers.
• Robustness is crucial for accurate image interpolation, especially preserving edge structures.

Purpose of the Study:

• To propose a novel image interpolation algorithm using regularized local linear regression (RLLR).
• To enhance robustness against outliers compared to OLS-based methods.
• To improve the preservation of image edge structures.

Main Methods:

• Replaced OLS error norm with moving least squares (MLS) error norm for robust local structure estimation.
• Incorporated an l(2)-norm penalty for estimator complexity to ensure stability and prevent overfitting.
• Integrated manifold structure using both measured and unmeasured data points, leveraging geometric properties for smoothness preservation.

Main Results:

• The proposed RLLR algorithm demonstrates robust performance in image interpolation.
• Experimental results show competitive performance against state-of-the-art methods.
• Significant improvement in preserving image edge structures was observed.

Conclusions:

• RLLR offers a robust and effective approach to image interpolation.
• The method successfully addresses the limitations of OLS in handling outliers.
• This technique shows promise for advanced image processing applications requiring high fidelity.