Spectral embedded hashing for scalable image retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Large-scale image retrieval demands efficient hashing methods.
• Existing spectral hashing methods struggle with nonlinear data manifolds and high dimensionality.
• Handling out-of-sample data is crucial for practical image retrieval systems.

Purpose of the Study:

• To propose a novel graph-based hashing method, spectral embedded hashing (SEH), for large-scale image retrieval.
• To extend SEH to kernel SEH (KSEH) for improved efficiency and effectiveness with high-dimensional data.
• To demonstrate that existing hashing methods can be considered special cases of KSEH.

Main Methods:

• Introduced a new regularizer in spectral hashing's objective function to manage mismatches between Hamming embedding and low-dimensional representations.
• Employed linear regression for out-of-sample data handling and to better manage data from nonlinear manifolds.
• Developed kernel SEH (KSEH) using nonlinear regression for high-dimensional data and an efficient eigenvalue decomposition solver.

Main Results:

• SEH effectively handles data sampled from nonlinear manifolds by controlling embedding-data representation mismatches.
• KSEH significantly improves efficiency and effectiveness for high-dimensional data in image retrieval.
• Comprehensive experiments on CIFAR, Tiny-580K, NUS-WIDE, and Caltech-256 datasets validated the proposed methods' effectiveness.

Conclusions:

• SEH and KSEH offer superior performance for large-scale image retrieval compared to existing methods.
• The proposed regularizer and kernel extension enhance robustness to data nonlinearity and dimensionality.
• The developed methods provide efficient and effective solutions for complex image retrieval tasks.