Enhancing Deep Hashing With GCN-Based Models for Efficient Similarity Search

Abstract

Deep hashing models are employed to efficiently store and swiftly search large-scale datasets where data are high dimensional. Their optimization of the loss function and the non-differentiable sign function can lead to inadequate backpropagation learning, resulting in hash codes that may be challenging for similarity preserving. Additionally, CNN-based deep hashing is not well-suited for capturing arbitrary relationships in non-grid structures. To address these issues, this study designs a deep hashing model using differentiable sign activation functions to enable effective backpropagation and proposes a GCN-based hashing approach suitable for non-grid structures. This remedy is able to design a deep hashing model to map high-dimensional data to low-dimensional hash code while it preserves data similarity and enables fast search. This approach shows improved accuracy of hash codes and faster similarity search compared to the studied CNN-based methods. The GCN layers leverage non-grid data structures to transform high-dimensional data into low-dimensional hashes while preserving similarity between data. Our approach advances deep hashing techniques by overcoming limitations of previous models and offers a promising solution for efficient similarity searches in complex data structures. Future work should focus on optimizing GCN-based hashing for large-scale datasets and validating the model across diverse applications.

Keywords

• Self-supervised deep hashing
• graph convolutional network
• feature embedding
• similarity preserving
• image retrieval