Local Semantic Correlation Modeling Over Graph Neural Networks for Deep Feature Embedding and Image Retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep feature embedding seeks to learn image representations that minimize intra-class and maximize inter-class distances.
• Current methods often rely on deep neural networks with specialized loss functions.

Purpose of the Study:

• To propose a new approach for deep feature embedding using graph neural networks.
• To leverage local correlation structures in the feature space for improved embeddings.

Main Methods:

• A graph neural network (GNN) was developed to model local image correlations.
• Graph edges predict correlation scores between neighboring images.
• Graph nodes learn to generate embedded features via weighted summation of neighbors' features.

Main Results:

• The proposed GNN method demonstrated superior performance in image retrieval tasks.
• Significant improvements were observed, particularly in top-1 recall rates.
• Outperformed existing state-of-the-art methods by a considerable margin.

Conclusions:

• The graph-based approach effectively captures and utilizes local image correlations for discriminative feature embedding.
• This method offers a promising advancement for image retrieval and related computer vision applications.