Unsupervised Deep Hashing With Pseudo Labels for Scalable Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep hashing (DH) methods are popular for efficient image similarity searching.
• Unsupervised DH algorithms often degrade in performance due to poor visual feature distribution capture.
• Label-free scenarios present a significant challenge for current unsupervised DH methods.

Purpose of the Study:

• To propose a novel unsupervised deep hashing framework to address performance degradation in label-free scenarios.
• To convert unsupervised DH models into supervised ones by discovering pseudo-labels.
• To preserve visual feature distribution by unifying likelihood maximization, mutual information maximization, and quantization error minimization.

Main Methods:

• Developed a framework to discover pseudo-labels for unsupervised deep hashing.
• Integrated likelihood maximization, mutual information maximization, and quantization error minimization.
• Evaluated the method on three benchmark datasets.

Main Results:

• The proposed method significantly improves performance over state-of-the-art unsupervised hashing algorithms.
• Pseudo-label discovery effectively enhances the unsupervised DH model.
• The unified optimization approach preserves visual feature distribution.

Conclusions:

• The novel framework effectively tackles the challenges of unsupervised deep hashing.
• The method demonstrates superior performance in image similarity searching tasks.
• This approach offers a promising direction for future unsupervised DH research.