Unsupervised Discrete Hashing With Affinity Similarity.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised hashing enhances image retrieval but requires extensive manual labeling.
• Unsupervised methods struggle with semantic information capture and quantization errors due to relaxed binary constraints.

Purpose of the Study:

• To develop an unsupervised hashing method that overcomes limitations of existing approaches for image retrieval.
• To improve semantic information capture and minimize quantization errors in unsupervised hashing.

Main Methods:

• Proposed a novel Unsupervised Discrete Hashing (UDH) method.
• Introduced a balanced graph-based semantic loss to capture data affinity.
• Developed an orthogonal consistent loss for semantic leveraging and code independence.
• Integrated discrete optimization to preserve binary constraints and reduce quantization errors.

Main Results:

• UDH effectively captures semantic information crucial for robust binary code learning.
• The method successfully imposes independence of codes while leveraging semantic information.
• Preserving discrete binary constraints significantly alleviates quantization errors.
• UDH demonstrates superior performance compared to state-of-the-art unsupervised methods in image retrieval tasks.

Conclusions:

• UDH offers an effective solution for unsupervised image retrieval by addressing semantic gap and quantization errors.
• The proposed method provides a viable alternative to supervised hashing, reducing the need for manual annotations.
• UDH advances the field of unsupervised hashing, paving the way for more efficient and scalable image retrieval systems.