On the performance of neuronal matching algorithms.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• The visual correspondence problem is crucial for many computer vision tasks.
• Existing algorithms like Self Organizing Map (SOM) and Dynamic Link Matching (DLM) are candidates for solving this problem.

Purpose of the Study:

• To modify the Self Organizing Map (SOM) for neighborhood- and feature-preserving mapping of image planes.
• To compare the convergence speed and scaling behavior of the modified SOM and Dynamic Link Matching (DLM) on a benchmark problem.

Main Methods:

• Modification of the Self Organizing Map (SOM) algorithm.
• Implementation and testing of both modified SOM and Dynamic Link Matching (DLM) on a benchmark dataset.
• Analysis of convergence speed and computational complexity with varying pattern sizes.

Main Results:

• The modified SOM requires a large number of simple update steps, while DLM requires fewer, more complex steps.
• Results indicate exponential scaling for SOM versus polynomial scaling for DLM as pattern size increases.
• A parameter adjustment rule for DLM applicable to all problem sizes was successfully developed, unlike for SOM.

Conclusions:

• Dynamic Link Matching (DLM) exhibits more favorable scaling properties than the modified Self Organizing Map (SOM) for the visual correspondence problem.
• The findings suggest DLM is a more efficient solution for larger-scale visual correspondence tasks.
• A generalizable parameter tuning strategy for DLM was established, enhancing its practical applicability.