Rotation-Covariant Texture Learning Using Steerable Riesz Wavelets.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Texture analysis is crucial for image understanding.
• Existing methods often struggle with variations in scale, orientation, and illumination.
• A robust and adaptable texture learning framework is needed.

Purpose of the Study:

• To develop a rotation-covariant texture signature learning framework.
• To enhance texture classification accuracy and robustness.
• To create a method that automatically learns optimal scales and directions.

Main Methods:

• Utilizing linear combinations of Riesz wavelets.
• Employing kernel support vector machines for texture signature learning.
• Iteratively optimizing local orientations for rotation covariance.

Main Results:

• Achieved high classification accuracies (97-98%) on benchmark datasets (Outex_TC_00010, Outex_TC_00012, Contrib_TC_00000).
• Demonstrated robustness to changes in image orientation and illumination.
• Observed rapid convergence of class-wise signatures, indicating effective feature space projection.

Conclusions:

• The proposed texture learning approach effectively exploits local organizations of scales and directions.
• The framework offers high performance and robustness for texture classification.
• It is expected to perform well across various computer vision applications.