Restored texture segmentation using Markov random fields.

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Texture segmentation is vital for image analysis but degraded by noise.
• Noisy texture segmentation is increasingly important for applications like automated inspection and medical imaging.

Purpose of the Study:

• To develop and evaluate a robust three-phase approach for segmenting textures contaminated with Gaussian and salt-and-pepper noise.
• To improve segmentation accuracy compared to existing benchmark methods.

Main Methods:

• A three-phase approach involving image restoration, Markov Random Fields (MRF), and objective customization of Median Filter.
• Restoration phase utilizes high-performing techniques from recent literature.
• Segmentation phase employs a novel MRF-based technique with a customized Median Filter.

Main Results:

• Achieved up to 16% accuracy improvement for salt-and-pepper noise (70% density) and 15.1% for Gaussian noise (variance 50) on Brodatz textures.
• Demonstrated 4.08% accuracy improvement for Gaussian noise (variance 10) and 2.47% for salt-and-pepper noise (20% density) on Prague textures.
• The proposed method shows significant enhancements over benchmark approaches.

Conclusions:

• The developed three-phase method effectively segments noisy textures, offering substantial accuracy improvements.
• The approach is applicable to diverse image analysis tasks including satellite imagery, medical imaging, and industrial inspection.