Semi-Supervised Image Retrieval Through Particle Competition and Cooperation Combined with Manifold Learning

• Vinicius Atsushi Sato Kawai UNESP
• Gustavo Rosseto Leticio UNESP
• Daniel Carlos Guimarães Pedronette UNESP

Resumo

The growing volume of unstructured visual data poses significant challenges for content-based image retrieval, particularly in scenarios where labeled information is scarce. This motivates the development of semi-supervised strategies capable of enhancing retrieval performance using limited supervision. In this work, we propose a novel approach that combines manifold learning and class-based probabilistic modeling for semi-supervised image retrieval. The method employs the Particle Competition and Cooperation (PCC) model to generate class dominance vectors using only a small portion of labeled data. These vectors are then fused with low-dimensional embeddings derived from Uniform Manifold Approximation and Projection (UMAP), yielding a representation that preserves neighborhood structure while capturing semantic class distributions. Experiments on three image datasets—Oxford 17 Flowers, Corel5k, and CUB-200—demonstrate the effectiveness of the proposed approach, which consistently outperforms baseline descriptors and isolated enhancements.

Palavras-chave: Training, Visualization, Sensitivity, Image retrieval, Pipelines, Semantics, Probabilistic logic, Transformers, Vectors, Manifold learning