Fast-learning adaptive-subspace self-organizing map: an application to saliency-based invariant image feature...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• The adaptive-subspace self-organizing map (ASSOM) is valuable for generating invariant features and data visualization.
• A significant limitation of ASSOM is its slow learning process.

Purpose of the Study:

• To propose two fast implementations of ASSOM to accelerate its learning procedure.
• To investigate the application of these improved ASSOM methods for saliency-based invariant feature construction in image classification.

Main Methods:

• Developed two novel ASSOM implementations by analyzing its basis rotation operator and objective function.
• Proposed two schemes for feature construction: single ASSOM cumulative activity map and multi-ASSOM joint cumulative activity map.
• Evaluated both schemes on the Corel photo database for image classification.

Main Results:

• The multi-ASSOM scheme, utilizing one ASSOM per image category, demonstrated superior performance in image classification.
• The enhanced ASSOM implementations significantly boosted learning speed.
• The multi-ASSOM scheme showed promising results when applied to adult image filtering.

Conclusions:

• Fast ASSOM implementations effectively address the slow learning issue.
• The multi-ASSOM scheme offers an advantageous approach for image classification and content filtering tasks.
• This research contributes to more efficient invariant feature generation and its practical applications.