An Adaptive Multi-Granularity Graph Representation of Image via Granular-ball Computing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph Neural Networks (GNNs) face challenges in deep structure learning and parameter management for comprehensive node feature extraction.
• In computer vision, GNNs often exhibit lower classification accuracy than Convolutional Neural Networks due to difficulties in image-to-graph representation.
• Existing methods struggle with adaptively generating graph structures and edges based on image semantics.

Purpose of the Study:

• To propose a novel method for converting images into graph representations that overcomes limitations of existing approaches.
• To enhance the performance of GNNs in vision tasks by improving graph construction.
• To enable adaptive generation of graph blocks and edges based on image semantics without manual annotation.

Main Methods:

• A novel method utilizing granular-ball computing to convert images into graphical forms.
• Dynamic generation of block nodes with varying sizes and corresponding edges.
• The approach does not require manual annotation or other learning methods for graph construction.

Main Results:

• The proposed method effectively captures semantic information within the generated graph representations.
• Significantly enhanced accuracy in vision tasks compared to state-of-the-art methods.
• Achieved superior performance with a reduced number of parameters.

Conclusions:

• The granular-ball computing approach offers a promising solution for image-to-graph conversion in GNNs.
• This method improves GNNs' effectiveness in vision tasks by enabling adaptive and semantically rich graph representations.
• The work has substantial implications for advancing GNN performance in computer vision applications.