Manuscript received June 14, 2024; revised July 29, 2024; accepted December 11, 2024; published July 17, 2025.

Abstract—The ImageNet dataset, which features a hierarchical structure based on the WordNet ontology, has been widely used for training and evaluating image classification models. However, researchers have not fully explored the potential benefits of leveraging this hierarchical structure for both image classification and retrieval tasks. This paper examines how incorporating hierarchical relationships between object categories during model training and inference can enhance image classification accuracy and retrieval performance. We propose a novel approach that integrates a hierarchical loss function and inference strategy to capture and utilize the semantic relationships encoded in the ImageNet hierarchy. Our method demonstrates improved classification accuracy compared to baseline models trained on a flattened version of ImageNet, highlighting the importance of hierarchical structure in the learning process. We show particular improvements for classes with limited training data, achieving accuracy increases of up to 3.2% for classes with fewer than 1000 samples. Additionally, we demonstrate how the hierarchical structure can be leveraged for efficient and semantically meaningful image retrieval. By utilizing the semantic relationships between categories, our approach enables more accurate and relevant retrieval results. The proposed techniques advance image classification and retrieval systems by harnessing the rich semantic information encoded in hierarchically structured datasets like ImageNet. Our findings emphasize the significance of incorporating hierarchical knowledge in visual recognition tasks while highlighting the trade-offs between semantic relevance and visual distinctiveness. This research paves the way for more effective and interpretable image classification and retrieval methods, particularly in scenarios with limited training data.

Keywords—image classification, image retrieval, ImageNet, hierarchical structure, limited training data

Cite: Luis E. Muñoz Guerrero, Yony F. Ceballos, and Luis D. Trejos Rojas, "Leveraging ImageNet's Hierarchical Structure for Enhanced Image Classification and Retrieval," Journal of Image and Graphics, Vol. 13, No. 4, pp. 304-314, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.