Variational region-based segmentation using multiple texture statistics.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Classical texture-based segmentation relies on pixel-level analysis, facing challenges with neighborhood size and feature dissimilarity.
• Existing methods often require experimental tuning of parameters, impacting performance and accuracy.

Purpose of the Study:

• To propose and evaluate a robust region-level variational criterion for supervised and unsupervised texture-based image segmentation.
• To demonstrate the effectiveness of the region-based formulation compared to common variational approaches.

Main Methods:

• Developed a region-level energy criterion using nonparametric distributions of filter responses for characterizing image regions.
• Implemented supervised segmentation via minimizing similarity between region statistics and texture prototypes, alongside a boundary-based functional.
• Utilized a level-set formulation for optimizing the variational criteria in both supervised and unsupervised scenarios.

Main Results:

• The proposed region-level methods effectively mitigate limitations associated with classical pixel-based texture segmentation.
• The generic similarity measure, based on Kullback-Leibler divergences, allows optimal fusion of diverse texture features.
• Evaluated effectiveness and robustness on Brodatz and natural images, showing advantages over active contour methods.

Conclusions:

• The region-level variational approach offers a more robust and effective solution for texture-based image segmentation.
• This formulation overcomes the critical dependency on neighborhood size and shape inherent in traditional methods.
• The method provides a unified framework for both supervised and unsupervised segmentation, adaptable to various texture features.