Domain Adaptation Preconceived Hashing for Unconstrained Visual Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Learning to hash is crucial for efficient image retrieval, but traditional methods fail with distribution gaps between query and retrieval data.
• Unconstrained visual cues (illumination, pose, background) create significant distribution discrepancies, degrading performance.

Purpose of the Study:

• To propose an efficient and transferable hashing model for unconstrained cross-domain visual retrieval.
• To develop an unsupervised hashing method that learns domain-invariant representations.

Main Methods:

• Domain Adaptation Preconceived Hashing (DAPH) is introduced for learning transferable hashing codes.
• A domain-invariant feature transformation is learned using marginal discrepancy distance minimization and feature reconstruction.
• A domain adaptation preconceived quantization loss is proposed to enhance retrieval discrimination.

Main Results:

• DAPH demonstrates effectiveness in learning domain-invariant hashing representations.
• The method preserves content while ensuring domain adaptability of hashing codes.
• Experiments show DAPH outperforms state-of-the-art methods in unconstrained single- and cross-domain retrieval.

Conclusions:

• DAPH offers a novel approach to unconstrained visual retrieval by integrating domain adaptation into hashing.
• The proposed method successfully addresses the distribution gap challenge in cross-domain image retrieval.
• DAPH provides a robust and efficient solution for real-world image retrieval scenarios.