Rotation-invariant texture retrieval via signature alignment based on steerable sub-Gaussian modeling

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Texture retrieval is crucial for image analysis and content-based image retrieval (CBIR).
• Existing methods often struggle with rotation invariance, requiring image rotation or complex feature recalculation.
• Developing efficient and robust rotation-invariant texture retrieval remains a significant challenge.

Purpose of the Study:

• To propose a novel and efficient rotation-invariant texture retrieval method.
• To leverage a steerable sub-Gaussian model for robust feature extraction.
• To achieve high retrieval accuracy with reduced computational complexity.

Main Methods:

• Construction of a steerable multivariate sub-Gaussian model to capture image properties and their rotated versions.
• Feature extraction via estimation of covariations between orientation subbands of a steerable pyramid.
• Development of rotation-invariant signatures and similarity measurement using a matrix-based norm with angular signature alignment.

Main Results:

• The proposed method demonstrates a lower average retrieval error compared to previous techniques with similar computational complexity.
• The system achieves competitive performance against state-of-the-art retrieval systems.
• Experimental results validate the effectiveness of the angular signature alignment for rotation invariance.

Conclusions:

• The developed retrieval method offers an effective balance between computational complexity and average retrieval performance.
• The steerable sub-Gaussian model provides a robust foundation for rotation-invariant texture analysis.
• This approach advances the field of content-based image retrieval by addressing rotation invariance efficiently.