Color Sparse Representations for Image Processing: Review, Models, and Prospects.

Area of Science:

• Image Processing
• Computer Vision
• Machine Learning

Background:

• Sparse representations are crucial for image processing, with extensions for handling multi-channel color images.
• Existing dictionary-learning models for color images vary in their approach to representing color information.

Purpose of the Study:

• To review and unify dictionary-learning-based sparse representation models for color images.
• To introduce a novel color filtering model for efficient and effective color image representation.

Main Methods:

• A unifying framework based on the degrees of freedom of linear filtering/transformation of color channels.
• Reformulation of the scalar quaternionic linear model to demonstrate equivalence with constrained matrix-based color filtering.
• Introduction of a new model using unconstrained filters where spatial information is encoded by atoms and colors by filters.

Main Results:

• Demonstrated the equivalence between the scalar quaternionic linear model and constrained matrix-based color filtering.
• The new model encodes spatial morphologies with atoms and colors with filters, efficiently handling color variability.
• Color variability is managed by color filters, not by increasing dictionary size, leading to efficient color representation.

Conclusions:

• The proposed framework provides a unified view of existing sparse representation models for color images.
• The novel color filtering model offers an efficient approach to color image representation by separating spatial and color information.
• This method enhances the representation of color variability without excessively increasing model complexity.