Texture classification by modeling joint distributions of local patterns with gaussian mixtures

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning

Background:

• Texture classification relies on analyzing local pixel patterns, often using joint probability density functions (jPDFs).
• Estimating jPDFs with joint histograms (jHSTs) is challenging due to data sparsity and computational costs.
• Existing methods like cooccurrence matrices and Local Binary Patterns (LBPs) reduce information loss but still result in significant data reduction.

Purpose of the Study:

• To introduce a supervised texture classification framework that minimizes information loss from jPDF estimation.
• To develop a method for efficient and accurate jPDF modeling and comparison for texture analysis.
• To achieve rotation invariance in texture descriptors.

Main Methods:

• Filtering local texture neighborhoods using a filter bank.
• Parametrically describing the jPDF of filter responses using Gaussian Mixture Models (GMMs).
• Computing distances between jPDFs efficiently in closed form from GMM parameters.
• Extending the descriptor for rotation invariance.

Main Results:

• GMM parameters can be reliably estimated from small image regions.
• The GMM-based framework significantly reduces information loss compared to traditional methods.
• Combining LBP difference filters with GMM outperforms classical LBP and its extensions.
• Using Wavelet Frame Transform (WFT) filters with the GMM framework surpasses spin image and RIFT descriptors on the Brodatz dataset.

Conclusions:

• The proposed GMM-based framework offers a robust and efficient approach to texture classification.
• This method effectively models complex texture patterns while mitigating information loss.
• The framework demonstrates superior performance and rotation invariance, advancing texture analysis techniques.