Texture classification and retrieval using shearlets and linear regression

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional wavelets struggle with image discontinuities like edges.
• Shearlets offer improved image characterization capabilities.
• Statistical modeling of wavelet subbands is common for image recognition.

Purpose of the Study:

• To develop novel texture classification and retrieval methods using shearlets.
• To model adjacent shearlet subband dependencies for enhanced image analysis.
• To overcome limitations of complex subband coefficients in texture representation.

Main Methods:

• Utilized shearlets, an extension of wavelets, for image characterization.
• Employed linear regression to model adjacent shearlet subband dependencies.
• Developed texture classification using two energy features per subband.
• Implemented texture retrieval with contourlet domain statistics and a pseudo-feedback mechanism.

Main Results:

• Proposed shearlet-based methods significantly outperform current state-of-the-art.
• Demonstrated superior performance in texture classification and retrieval tasks.
• Validation on five large texture datasets confirmed effectiveness.

Conclusions:

• Shearlet modeling of subband dependencies offers a powerful approach for texture analysis.
• The proposed methods provide a robust solution for image recognition and retrieval.
• This work advances the field of image processing with novel shearlet applications.