Generalized Coupled Dictionary Learning Approach With Applications to Cross-Modal Matching.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Coupled dictionary learning (CDL) is a powerful technique for various applications.
• Existing CDL methods struggle with unpaired data and imbalanced class distributions across modalities.
• Cross-modal matching requires robust methods for handling data from different sources.

Purpose of the Study:

• To extend coupled dictionary learning (CDL) for improved cross-modal matching.
• To develop a CDL approach that can handle both paired and unpaired data effectively.
• To enhance CDL for better performance in cross-modal retrieval and classification tasks.

Main Methods:

• Learned two dictionaries, one for each modality, enabling sparse data representation.
• Transformed sparse coefficients to maximize correlation for same-class data and minimize for different-class data.
• Incorporated a discriminative coupling term to improve classification capabilities.

Main Results:

• The generalized CDL approach demonstrated superior performance on multiple public cross-modal datasets.
• The method achieved state-of-the-art results for both paired and unpaired cross-modal matching scenarios.
• Experiments confirmed the effectiveness of the approach on datasets like CUHK photosketch and HFB.

Conclusions:

• The proposed generalized CDL method is highly effective for cross-modal matching with paired and unpaired data.
• The approach offers a significant improvement over existing methods for cross-modal retrieval and classification.
• This work provides a robust framework for handling heterogeneous data in machine learning.