Deep Semisupervised Class- and Correlation-Collapsed Cross-View Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Multiview data presents challenges in computer vision due to distribution inconsistencies.
• Existing methods often focus on two-view problems, limiting scalability.
• Exploiting both labeled and unlabeled data is crucial for enhancing multiview relationships.

Purpose of the Study:

• To propose a novel deep semi-supervised framework for cross-view learning.
• To address the limitations of existing methods by handling more than two views (U views).
• To improve cross-view retrieval and classification by learning a discriminative shared space.

Main Methods:

• Developed deep semi-supervised classes- and correlation-collapsed cross-view learning (DSC3L).
• Learned U view-specific deep transformations to project multiple views into a shared space.
• Integrated supervised and unsupervised learning by minimizing Kullback-Leibler (KL) divergences for class and correlation collapsing.

Main Results:

• The proposed DSC3L framework effectively handles U-view multiview data (U ≥ 2).
• Achieved improved performance in cross-view retrieval and classification tasks.
• Demonstrated effectiveness across five widely used public datasets.

Conclusions:

• DSC3L provides a robust solution for leveraging multiview data in computer vision.
• The framework successfully bridges the heterogeneous gap between different views.
• The method offers a significant advancement in semi-supervised cross-view learning.