Multimodal Image Classification by Multiview Latent Pattern Extraction, Selection, and Correlation

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• The big data era offers opportunities to integrate heterogeneous data sources.
• Multiview learning excels at extracting complementary information from multiple data modalities.

Purpose of the Study:

• Propose a novel framework, Multiview Latent Space Projection (MVLSP), for discriminative feature integration.
• Facilitate binary and multiclass classifications using heterogeneous data sources.

Main Methods:

• MVLSP maps features from multiple views into a common latent space.
• The framework extends to more than two views via view-by-view matching.
• Feature selection is achieved by incorporating a class view for feature-label correlation.
• Optimizes integration of latent patterns based on their correlations.

Main Results:

• Demonstrated effectiveness on the prostate image dataset.
• Successfully integrated features from multiple heterogeneous sources.
• Achieved improved classification performance.

Conclusions:

• MVLSP provides a scalable and effective approach for multiview learning.
• The proposed method enhances feature integration and selection for classification.
• Highlights the potential of latent space projection in big data analysis.