Discriminative Structured Dictionary Learning on Grassmann Manifolds and Its Application on Image Restoration

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Image restoration is crucial for various imaging applications but faces challenges in capturing geometric structures.
• Existing sparse representation methods struggle with representing local image structures accurately.

Purpose of the Study:

• To propose a new sparse representation formulation with block-orthogonal constraints for enhanced image restoration.
• To improve the accurate representation of local image structures.

Main Methods:

• Developed a discriminative structured dictionary learning framework with manifold structures.
• Implemented an alternating minimization scheme for dictionary block structure updates on the Grassmann manifold and automatic atom sparsification.
• Utilized Riemannian conjugate gradient for efficient local subspace tracking with convergence guarantees.

Main Results:

• The proposed method demonstrates superior performance in removing mixed Gaussian-impulse noise compared to state-of-the-art techniques.
• Experiments on diverse datasets validate the effectiveness of the new approach.

Conclusions:

• The block-orthogonal constraint in sparse representation offers a robust solution for image restoration.
• The proposed dictionary learning and optimization scheme effectively handles local image structures and noise reduction.