Graph Convolutional Network Hashing.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Neural Networks

Background:

• Graph-based hashing methods are vital for large-scale image retrieval.
• Existing methods often rely on computationally expensive binary quadratic programs, limiting scalability.
• There's a need for efficient and scalable graph hashing techniques.

Purpose of the Study:

• To propose a novel graph convolutional network-based hashing framework (GCNH) for scalable image retrieval.
• To address the limitations of traditional graph hashing methods regarding scalability and out-of-sample extension.
• To demonstrate the effectiveness of GCNH in semi-supervised hashing scenarios.

Main Methods:

• Developed a graph convolutional network-based hashing framework (GCNH).
• Introduced an asymmetric graph convolutional (AGC) layer for efficient spectral convolution.
• Applied GCNH to image sets and affinity graphs for similarity-preserving binary embedding.

Main Results:

• GCNH demonstrates superior scalability compared to traditional methods.
• The framework effectively handles out-of-sample extension.
• Consistent advantages observed in semi-supervised hashing evaluations on CIFAR-10, NUS-WIDE, and ImageNet datasets.

Conclusions:

• GCNH offers a scalable and effective solution for graph-based hashing in image retrieval.
• The asymmetric graph convolutional layer is key to GCNH's performance.
• GCNH shows significant promise for semi-supervised learning tasks with limited labeled data.