A Comparison of Feature Extraction Models for Images with Multiple Annotations

An experimental comparison of the accuracy of neural networks for feature extraction from images has been carried out. A classification is chosen as an experimental task for the dataset with multiple annotations, that was prepared in advance. The existing architectures and approaches to training foundation models for feature extraction are analyzed theoretically. The results of the experimental comparison of the models under consideration are analyzed in detail paying extra attention to the imbalanced classes. The reasons for the decrease in recognition accuracy in experiments are investigated, and the most erroneous classes are identified. The analysis of spatial relationships between feature vectors has been carried out to assess the applicability of metric classification algorithms. The latent spaces of the considerable models are visualized and the comparison of the shape of the resulting clusters is provided. As a result, open-source models that can be used in practical tasks without fine-tuning are identified and the most accurate and accessible one is chosen.