Unsupervised Semantic-Preserving Adversarial Hashing for Image Search

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing is crucial for large-scale image retrieval, with deep learning methods showing promise.
• Existing deep hashing methods often require large labeled datasets, limiting practical application.
• Unsupervised learning offers a potential solution to overcome data labeling challenges.

Purpose of the Study:

• To develop an efficient unsupervised deep hashing method for large-scale image retrieval.
• To leverage semantic similarity in training data for improved hash code generation.
• To address the limitations of supervised deep hashing methods by eliminating the need for labeled data.

Main Methods:

• A generative adversarial framework comprising encoder, generative, and discriminative networks was designed.
• An encoder network learns hash codes from images, while a generative network creates images from hash codes.
• A novel semantic similarity matrix, incorporating feature and neighbor similarities, guides the adversarial learning process.

Main Results:

• The adversarial training successfully created coherent encoder and generative networks for efficient hash code output.
• The semantic similarity matrix effectively preserved data semantics in Hamming space during unsupervised learning.
• Experimental results demonstrated superior performance compared to state-of-the-art unsupervised hashing methods.

Conclusions:

• The proposed unsupervised generative adversarial hashing framework is effective for large-scale image retrieval.
• The method achieves performance comparable to supervised approaches without requiring labeled data.
• This approach offers a viable alternative for real-world image retrieval applications with limited labeled data.