SCH-GAN: Semi-Supervised Cross-Modal Hashing by Generative Adversarial Network

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-modal hashing enables retrieval across heterogeneous data types by mapping them to a common space.
• Supervised methods require large labeled datasets, which are difficult to acquire for multimedia data.
• Unlabeled data holds valuable information for modeling inter-modal correlations.

Purpose of the Study:

• To propose a novel semi-supervised cross-modal hashing approach (SCH-GAN) to address limitations of existing methods.
• To leverage generative adversarial networks (GANs) and reinforcement learning for improved cross-modal retrieval.
• To utilize unlabeled data effectively for enhanced hashing performance.

Main Methods:

• A generative adversarial network (GAN) framework where a generator selects margin examples and a discriminator distinguishes them from true positives.
• A reinforcement learning-based algorithm to train the SCH-GAN, using correlation scores as rewards for the generator.
• Training involves a minimax game between the generator and discriminator to optimize hashing.

Main Results:

• The proposed SCH-GAN demonstrates significant effectiveness in cross-modal retrieval tasks.
• Experimental results show superior performance compared to nine state-of-the-art methods on three benchmark datasets.
• The approach successfully utilizes unlabeled data to improve hashing accuracy.

Conclusions:

• SCH-GAN offers an effective semi-supervised solution for cross-modal hashing.
• The method overcomes the reliance on large labeled datasets by incorporating unlabeled data.
• This approach advances fast and flexible cross-modal retrieval capabilities.