Model Optimization Boosting Framework for Linear Model Hash Learning.

Area of Science:

• Computer Science
• Machine Learning
• Data Storage and Retrieval

Background:

• Efficient hashing is crucial for managing high-dimensional data like images and videos.
• Linear models are common in hashing due to their efficiency, but often require complex objective functions for better accuracy.
• Existing methods focus on adding constraints or penalties to linear models to capture data characteristics.

Purpose of the Study:

• To introduce a novel self-improvement framework, Model Boost (MoBoost), for linear-based hashing methods.
• To enhance model parameter optimization without introducing additional constraints or penalty terms.
• To improve the accuracy and stability of existing linear hashing techniques.

Main Methods:

• The MoBoost framework repeatedly applies a linear hashing method to generate multiple sets of hash codes for training samples.
• Two novel fusion strategies are employed to combine these multiple hash codes into a single, unified set.
• Two new criteria are proposed to evaluate the quality of hash bits during the fusion process.
• New parameters for the linear hash function are learned based on the fused hash codes.

Main Results:

• The proposed MoBoost framework significantly improves the accuracy of linear-based hashing methods.
• MoBoost achieves more precise and stable performance compared to original hashing methods.
• The framework incurs negligible additional time and space costs.
• Extensive experiments on four benchmark datasets validate the superior performance of MoBoost.

Conclusions:

• MoBoost offers a generalizable framework that can be adopted by various existing linear-based hashing methods.
• The framework effectively enhances model parameter optimization for improved hashing performance.
• MoBoost provides a cost-effective solution for achieving higher accuracy and stability in high-dimensional data retrieval.