Discrete Spectral Hashing for Efficient Similarity Retrieval

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• High-dimensional data analysis requires efficient representation techniques.
• Hashing methods learn binary representations for data organization and storage.
• Unsupervised spectral hashing is effective for manifold embedding but faces challenges.

Purpose of the Study:

• To address limitations in existing spectral hashing methods, specifically inefficient spectral candidates and complex binary constraints.
• To propose novel spectral hashing techniques that enhance both spectral solutions and discrete code efficiency.
• To develop efficient optimization algorithms for these new hashing methods.

Main Methods:

• Introduced Spectral Hashing with Spectral Rotation to find better spectral solutions.
• Developed Alternating Discrete Spectral Hashing using alternating projection for code constraints.
• Combined spectral rotation with spectral objectives in Discrete Spectral Hashing for simultaneous improvements.
• Provided efficient optimization algorithms with comparable time complexity.

Main Results:

• The proposed three methods, including Discrete Spectral Hashing, demonstrated significant improvements.
• Extensive experiments on four large-scale datasets for image retrieval were conducted.
• The novel methods outperformed several state-of-the-art spectral hashing techniques across various metrics.

Conclusions:

• The proposed spectral hashing methods effectively address existing challenges in unsupervised spectral hashing.
• These methods offer superior performance in high-dimensional data representation and image retrieval.
• The developed techniques provide efficient solutions for large-scale data analysis and storage demands.