Graph Convolutional Network Discrete Hashing for Cross-Modal Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks have advanced cross-modal hashing, but existing methods struggle with asymmetrical information between modalities (e.g., images vs. text).
• Current approaches often overlook leveraging information-rich modalities to support information-poor ones, leading to suboptimal performance.
• Previous methods incur quantization losses due to relaxed learning of hash functions.

Purpose of the Study:

• To propose a novel cross-modal hashing method that effectively bridges the information gap between different data types.
• To enhance feature representations by utilizing information-rich modalities to support information-poor ones.
• To achieve accurate discrete binary codes without relaxation-induced quantization losses.

Main Methods:

• Developed a graph convolutional network (GCN) discrete hashing (GCDH) method.
• Utilized GCN to represent labels as word embeddings, treating them as interdependent object classifiers.
• Employed an efficient discrete optimization strategy for learning binary codes without relaxation.

Main Results:

• The proposed GCDH method successfully bridges the information gap between modalities.
• Enhanced feature representations across modalities by leveraging label embeddings.
• Achieved superior performance compared to state-of-the-art methods on three benchmark datasets.

Conclusions:

• GCDH effectively addresses the information asymmetry challenge in cross-modal hashing.
• The discrete optimization strategy avoids quantization losses, leading to more accurate hash codes.
• This approach offers a significant improvement for cross-modal retrieval tasks.