Spatial general autoregressive model-based image interpolation accommodates arbitrary scale factors.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image interpolation is crucial for resizing digital images.
• Existing methods often struggle with arbitrary (non-integer) scaling factors.
• High-fidelity interpolation requires sophisticated modeling of image structures.

Purpose of the Study:

• To develop a novel image interpolation algorithm capable of handling arbitrary upscaling factors.
• To improve the modeling of image patterns for more accurate interpolation.
• To reduce visual artifacts commonly seen in image resizing.

Main Methods:

• Utilized a spatial general autoregressive model adapted for non-integer scale factors.
• Employed a gradient angle guided extension to incorporate more neighboring pixels for parameter estimation.
• Applied elastic network regularization to prevent overfitting and ensure accurate parameter estimation.
• Introduced an iterative curvature method for refining interpolation in high-variance image blocks.

Main Results:

• The proposed algorithm effectively suppresses visual artifacts across various integer and non-integer scaling factors.
• Demonstrated superior performance in objective and subjective image quality measures compared to existing methods.
• Successfully implemented a flexible image interpolation method for arbitrary scale factors.

Conclusions:

• The novel spatial general autoregressive model provides a robust framework for arbitrary image interpolation.
• The integration of gradient guidance, elastic net regularization, and iterative curvature refinement enhances interpolation accuracy and visual quality.
• This method offers a flexible and effective solution for image resizing challenges.