Cartoon-texture image decomposition using blockwise low-rank texture characterization

Area of Science:

• Computer Vision
• Image Processing
• Optimization

Background:

• Image decomposition is crucial for analyzing visual content.
• Existing methods struggle with complex texture patterns and image degradations.
• Textures exhibit local pattern repetition, a property not fully exploited.

Purpose of the Study:

• To develop a novel image decomposition technique separating cartoon and texture components.
• To introduce a new convex prior, the block nuclear norm (BNN), for texture characterization.
• To create a robust model handling various image degradations.

Main Methods:

• Proposed a new convex prior: the block nuclear norm (BNN).
• Formulated a convex optimization problem minimizing total variation and BNN.
• Employed variable splitting and the alternating direction method of multipliers for efficient computation.
• Developed a model capable of extracting texture patterns in different directions.

Main Results:

• The proposed method effectively decomposes images into cartoon and texture components.
• The block nuclear norm accurately characterizes texture based on local pattern repetition.
• The model demonstrates high selectivity for texture patterns.
• Achieved superior performance compared to state-of-the-art decomposition models.

Conclusions:

• The novel image decomposition technique offers significant improvements in separating image components.
• The block nuclear norm provides an effective prior for texture modeling.
• The method's robustness to degradations and pattern extraction capabilities enhance its applicability.