Performance evaluation of finite normal mixture model-based image segmentation techniques.

Area of Science:

• Computer Vision
• Statistical Modeling
• Image Processing

Background:

• Finite Normal Mixture (FNM) models are widely used for image segmentation.
• Current techniques follow a detection-estimation-classification paradigm.
• Performance evaluation of these methods is crucial for reliable image analysis.

Purpose of the Study:

• To develop a theoretical framework for evaluating FNM model-based image segmentation performance.
• To define and derive metrics for detection, estimation, and classification accuracy.
• To provide a method for assessing both theoretical limits and practical performance.

Main Methods:

• Defining probabilities of over-detection and under-detection for region detection.
• Deriving Cramer-Rao bounds for the variance of Expectation-Maximization (EM) and Classification-Maximization (CM) algorithm estimates.
• Defining misclassification probability for Bayesian classifiers and deriving a formula for segmentation error evaluation.

Main Results:

• Probabilities of over- and under-detection are derived in terms of model parameters and image quality.
• EM and CM algorithms yield asymptotically unbiased ML estimates under no-overlap conditions.
• A formula for evaluating segmentation errors is derived based on parameter estimates and classified data.

Conclusions:

• The proposed evaluation method offers theoretically approachable accuracy limits and practically achievable performance metrics.
• Theoretical and experimental results show good agreement.
• For moderate quality images, FNM-based segmentation demonstrates robust detection, accurate parameter estimates, and small segmentation errors.