Image search, Feature extraction, Dimensionality reduction, Multidimensional indexing, Information retrieval
Resumo
As image databases grow and evolve, the need for effective retrieval engines becomes increasingly critical. While Content-Based Image Retrieval (CBIR) has advanced considerably across various domains, its application in microbiology remains limited; particularly from a macroscopic perspective, where pure culture images often appear visually similar and are difficult to distinguish. This study pioneers a preliminary investigation into the development of a CBIR framework for accessing images of microorganisms grown on solid media. Our proposal leverages deep neural networks to extract image features and applies Uniform Manifold Approximation and Projection (UMAP) to reduce the dimensionality of feature vectors without sacrificing performance. To enable fast and efficient retrieval, we implement a K-dimensional tree structure in combination with the K-nearest neighbors algorithm. We evaluate the framework’s performance using a leave-one-out strategy, considering precision and recall, measured by mean average precision, along with execution time. Experiments on two original microbial image sets yield promising results, with vision transformers outperforming other pre-trained models. Our findings highlight the potential of the developed framework for accurately retrieving culture images and demonstrate its applicability in both clinical and microbiological research.
