Optimized Graph Learning Using Partial Tags and Multiple Features for Image and Video Annotation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multimedia annotation often faces challenges due to time constraints and manual tagging efforts.
• Supervised learning models benefit from utilizing both tagged and untagged data when labeled training data is scarce.
• Existing graph-based learning algorithms rely on similarity graphs, but their construction is often empirical and feature-dependent, neglecting label information.

Purpose of the Study:

• To propose a novel semi-supervised annotation approach that learns an optimized graph (OGL) from multi-cues.
• To enhance the accuracy of embedding relationships among data points by integrating partial tags and multiple features.
• To extend the OGL model to address the out-of-sample issue inherent in transductive methods.

Main Methods:

• Developed an optimized graph learning (OGL) framework for semi-supervised multimedia annotation.
• Incorporated multi-cues, including partial tags and multiple features, for robust graph construction.
• Extended the transductive OGL model to handle novel data points, addressing the out-of-sample problem.

Main Results:

• The proposed OGL method demonstrated superior performance in embedding data point relationships compared to existing approaches.
• Extensive experiments on image and video annotation tasks showed consistent improvements.
• The extended OGL model effectively resolved the out-of-sample issue, enhancing its practical applicability.

Conclusions:

• The OGL approach offers a more accurate and effective method for semi-supervised multimedia annotation.
• Integrating multiple features and partial labels in graph construction significantly boosts performance.
• The developed model provides a robust solution for both existing and novel data points in annotation tasks.