A Graph-Embedding Approach to Hierarchical Visual Word Mergence.

Area of Science:

• Computer Science
• Machine Learning
• Computer Vision

Background:

• Hierarchical merging of visual words is crucial for dimension reduction in bag-of-visual-words models for image classification.
• Existing supervised methods lack a unified approach and vary in formulation.
• A determined merging hierarchy is desirable for consistent results.

Purpose of the Study:

• To propose a unified hierarchical visual word merging approach using a graph-embedding framework.
• To develop a flexible method that accommodates preferred and undesired structures for word merging.
• To maintain computational efficiency and state-of-the-art merging speed.

Main Methods:

• A novel graph-embedding framework for unified hierarchical visual word merging.
• Integration of a fast search strategy for computational efficiency.
• Defining preferred and undesired structures to guide the merging process.

Main Results:

• The proposed approach achieves excellent image classification performance post-dimension reduction.
• It outperforms existing comparable visual word-merging methods.
• The method demonstrates flexibility in handling diverse merging requirements.

Conclusions:

• The unified graph-embedding approach offers a flexible and effective solution for hierarchical visual word merging.
• This work provides an open platform for developing and evaluating new merging criteria.
• The method maintains high classification accuracy with significant dimension reduction.