Common Visual Pattern Discovery via Directed Graph.

Area of Science:

• Computer Vision
• Graph Theory
• Image Analysis

Background:

• Object identification in image comparison is crucial for various applications.
• Existing methods often use undirected graphs, limiting directional information.
• Directed graphs offer superior modeling capabilities due to oriented link weights.

Purpose of the Study:

• To propose a directed graph (digraph) approach for identifying common visual objects in two images.
• To address the challenges of computing link weights and extracting subgraphs in digraphs.
• To demonstrate the superiority of the digraph approach over undirected graph methods.

Main Methods:

• Developed a novel n-ranking process and Gaussian link-weight mapping function to map undirected graphs to digraphs.
• Utilized an evolutionary iterative process based on non-cooperative game theory for subgraph extraction.
• Applied these methods for each scale-change factor to identify common visual patterns.
• Employed a topological splitting method for differentiating extracted patterns.

Main Results:

• The proposed digraph approach effectively computes directed link weights, considering neighbor influences and offering enhancements.
• Subgraph extraction is efficiently handled using game theory for non-symmetric matrices.
• The method successfully identifies common visual objects across different scale-change factors.
• Extensive simulations confirm the superior performance of the digraph approach compared to undirected graph methods.

Conclusions:

• The directed graph approach provides a robust framework for visual object identification in image comparison.
• The developed methods for link weight computation and subgraph extraction effectively address key challenges.
• This approach offers significant performance improvements over traditional undirected graph methods.