Multi-label guided graph attention network for education image retrieval

In recent years, deep learning has achieved remarkable success thanks to advanced neural network architectures and large-scale datasets manually labeled by humans. However, accurately and efficiently labeling large datasets is often costly and challenging, particularly in fields requiring specialized labeling expertise, such as healthcare. In this context, building a model capable of large-scale image retrieval without extensive manual labeling is a crucial need. This study proposes a multi-label learning method based on attentive graph convolutions called GLGM (Graph network combined with Local and Global features based on Multi-label techniques) to address the issue of detailed classification with coarsely labeled datasets. Specifically, within the framework of contrastive learning, our method generates labels interconnected through graph convolutions. Unlike self-supervised contrastive learning methods that link global and local image features to create a graph that represents specific object characteristics, GLGM introduces a common search space that supports image retrieval in the educational field and image retrieval in general based on advanced sample distance search algorithms. We demonstrate that the GLGM method can encompass many state-of-the-art approaches as special cases. Experiments show that GLGM achieves significant improvements over existing advanced methods on various datasets, including CIFAR-10 and MLIC-Edu (a dataset we collected ourselves for the educational image retrieval domain).