A New Method of Probability Density Estimation with Application to Mutual Information Based Image Registration.

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Estimating probability density of image intensities is crucial for image analysis.
• Traditional methods like histograms and mixture models have limitations, including sensitivity to parameter tuning and binning issues.

Purpose of the Study:

• To develop a robust, computationally efficient method for image intensity probability density estimation.
• To address limitations of existing sample-based density estimation techniques.
• To extend the method for joint density estimation between multiple images.

Main Methods:

• Utilizes a continuous image representation and relates probability density to image gradients along level sets.
• Exploits the geometry of the image surface and the ordering of intensity values.
• Avoids histogram binning and critical parameter tuning.

Main Results:

• The proposed method provides a robust and computationally efficient approach to probability density estimation.
• Demonstrates successful application in affine registration of 2D images using mutual information, even under high noise conditions.

Conclusions:

• The new density estimation technique offers advantages over traditional methods.
• It shows promise for applications in image registration and other image analysis tasks requiring accurate density estimation.