<p>The rapidly expanding volume of digital educational resources necessitates effective and efficient image retrieval systems. However, many contemporary deep hashing approaches for image retrieval may not fully capture the rich semantic information inherent in educational images or adequately leverage the complex relationships between associated labels, which is crucial for retrieval accuracy in this domain. This paper introduces G-hash, a novel framework for educational image retrieval that integrates Graph Attention Networks (GAT) with deep hashing techniques. G-hash employs Transformer encoders to extract robust global features from both query images and textual labels within the educational dataset. A key innovation is the use of a GAT module to explicitly model and explore the semantic correlations and co-occurrence patterns among labels, thereby generating more discriminative and context-aware label representations. These image and enhanced label features are then processed by a deep hashing module, which includes a hash layer with ‘tanh‘ and ‘sign‘ functions, to learn compact binary codes. This end-to-end architecture aims to generate hash codes that preserve semantic similarity while ensuring retrieval efficiency, significantly improving the ability to retrieve relevant educational images. The proposed G-hash model is poised to enhance the precision and speed of image search in educational applications.</p>