A Spatially Constrained Probabilistic Model for Robust Image Segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Hidden Markov Random Field (HMRF) models are used for probabilistic image segmentation.
• Existing HMRF models inadequately utilize spatial information for parameter estimation and class label assignment.
• Current methods exhibit sub-optimal performance, particularly for pixels at class boundaries.

Purpose of the Study:

• To develop a novel clique potential function and class label distribution for HMRF-based image segmentation.
• To enhance parameter estimation and class label distribution by incorporating image class parameters.
• To introduce an adaptive scaling parameter for dynamically measuring spatial information's contribution.

Main Methods:

• Developed a new clique potential function and class label distribution within the HMRF framework.
• Introduced an adaptive scaling parameter to modulate spatial information's influence on pixel class label estimation.
• Modified existing HMRF-based segmentation methods to integrate the proposed framework.

Main Results:

• The proposed framework significantly improves segmentation performance compared to existing HMRF models.
• Adaptive spatial information incorporation reduces misclassification at image class boundaries.
• Demonstrated effectiveness across diverse datasets including brain MR images, HEp-2 cells, and natural images.

Conclusions:

• The novel HMRF class label distribution enhances segmentation accuracy by adaptively utilizing spatial information.
• The proposed method offers improved performance irrespective of underlying image intensity distributions.
• This approach provides a more robust and accurate image segmentation technique for various applications.