FCPN: Pruning redundant part-whole relations for more streamlined pattern parsing

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Pattern parsing methods often create over-generalizations and redundant representations by combining diverse correlations.
• Existing methods struggle with detailed semantic boundary definition in complex segmentation tasks.

Purpose of the Study:

• To streamline pattern parsing and enhance segmentation accuracy.
• To introduce a novel network architecture for eliminating redundant associations in pattern parsing.

Main Methods:

• Proposed a Redundant Association Elimination Network (RAEN).
• Incorporated Capsule Attention Twisters (CATs) to refine part-whole relationships.
• Utilized Capsule-Attention Routing Agreement (CARA) to prevent redundant voting signals.

Main Results:

• RAEN effectively trims weak and interchangeable relationships between parts and wholes.
• Senior entities are updated only by primary entities meeting specific diversity and cohesiveness criteria.
• CARA protects against unnecessary voting signals, improving overall performance.

Conclusions:

• RAEN demonstrates superior performance compared to existing methods in facial and human segmentation.
• The proposed network excels at defining detailed semantic boundaries, offering a significant advancement in pattern parsing.