Cycle-Consistent Deep Generative Hashing for Cross-Modal Retrieval.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Cross-modal retrieval is challenging due to the lack of paired training data.
• Existing methods struggle to learn effective hash functions without direct supervision.

Purpose of the Study:

• To propose a novel deep generative approach for cross-modal retrieval.
• To learn hash functions effectively in the absence of paired training samples.
• To enhance semantic correlation between cross-modal data.

Main Methods:

• Utilizing a cycle consistency loss within an adversarial training scheme.
• Employing a generative adversarial network (GAN) to learn hash functions.
• Jointly optimizing hash functions and generative models for cross-modal embedding.

Main Results:

• The proposed method successfully learns hash functions without paired data.
• Cycle consistency loss strengthens the semantic correlation between modalities.
• Achieved state-of-the-art performance on large-scale cross-modal datasets.

Conclusions:

• The deep generative approach with cycle consistency loss is effective for unsupervised cross-modal retrieval.
• This method offers a robust solution for learning semantic hash codes across modalities.
• The generative nature minimizes information loss while maximizing cross-modal correlation.