Antipodally Invariant Metrics for Fast Regression-Based Super-Resolution

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Dictionary-based super-resolution (SR) algorithms rely on distance metrics for selecting dictionary atoms.
• The impact of metric choice on SR performance is often overlooked, with Euclidean distance being prevalent.
• Features in SR commonly reside on the unitary hypersphere, exhibiting antipodal properties.

Purpose of the Study:

• To investigate the influence of distance metrics on SR performance.
• To develop a faster and more accurate SR algorithm by addressing limitations of Euclidean distance.
• To introduce an effective antipodally invariant transform for SR.

Main Methods:

• A fast regression-based algorithm using anchored neighborhoods and sublinear search structures.
• Analysis of SR features on the unitary hypersphere and the concept of antipodes.
• Development and integration of an antipodally invariant transform into Euclidean distance calculations.
• Modification of the spherical hashing algorithm into an antipodally invariant spherical hashing scheme.
• A novel feature transform incorporating iterative backprojection for improved coarse approximation.

Main Results:

• The proposed antipodally invariant SR method achieves superior Peak Signal to Noise Ratio (PSNR) compared to state-of-the-art methods.
• The algorithm demonstrates significant speed improvements over existing SR techniques.
• The antipodally invariant transform enhances the selection of dictionary atoms, leading to better SR outcomes.
• The modified spherical hashing scheme matches the performance of pure antipodally invariant metrics.

Conclusions:

• The developed antipodally invariant SR algorithm offers a significant advancement in image super-resolution.
• The integration of antipodally invariant metrics provides a more optimal approach for SR feature selection.
• The method presents a compelling balance of improved image quality and computational efficiency.