Performance evaluations of correlations of digital images using different separability measures

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Accurate image registration is crucial for analyzing and comparing images from different viewpoints.
• Selecting appropriate reference images is a key challenge in two-dimensional image registration.
• Various statistical measures can quantify image separability for registration.

Purpose of the Study:

• To compare the effectiveness of four separability measures in selecting reference images for 2D image registration.
• To evaluate the performance of these measures using both real and synthetic aerial images.
• To assess the impact of different image views (down-looking vs. target-looking) on registration accuracy.

Main Methods:

• Employed four separability measures: Bayes probability of error, Chernoff bound, Bhattacharyya bound, and Fisher's criteria.
• Utilized area and edge correlations for image comparison.
• Conducted experiments on real and synthetic aerial images with down-looking and target-looking views.

Main Results:

• For real aerial images, target-looking views yielded better registration results with both area and edge correlations.
• For synthetic aerial images, down-looking views outperformed target-looking views for both correlation types.
• Synthetic images exhibited greater variation between views compared to real images.

Conclusions:

• The choice of image view significantly impacts registration performance, depending on whether images are real or synthetic.
• Separability measures are effective in guiding the selection of optimal reference images for image registration.
• Understanding view-dependent variations is essential for robust image registration systems.