Bidirectional Discrete Matrix Factorization Hashing for Image Search.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Unsupervised image hashing faces challenges due to limited supervision data.
• Existing methods often fail to preserve intrinsic visual data information, limiting performance.
• Reliance on large affinity matrices can be computationally intensive and insufficient.

Purpose of the Study:

• To propose a novel unsupervised image hashing method, bidirectional discrete matrix factorization hashing (BDMFH).
• To enhance the preservation of intrinsic data structure within learned binary codes.
• To achieve superior performance and computational efficiency compared to state-of-the-art methods.

Main Methods:

• BDMFH employs alternating processes for learning binary codes and recovering data.
• An inverse factorization model is designed to ensure learned binary codes retain original data structure.
• An efficient discrete optimization algorithm is developed for BDMFH.

Main Results:

• BDMFH significantly outperforms existing unsupervised image hashing techniques.
• The method demonstrates superior performance across three large-scale benchmark datasets.
• BDMFH achieves satisfactory computational efficiency.

Conclusions:

• BDMFH offers an effective solution for unsupervised image hashing by preserving intrinsic data structure.
• The proposed method achieves state-of-the-art performance with improved efficiency.
• BDMFH represents a significant advancement in the field of unsupervised hashing.