Resumen
The article deals with the stages of solving image clustering problems using pre-trained neural networks. Some composite solutions of the clustering problem are presented, where clustering methods are used at the last stage, and most of the work is the extraction of features and their preprocessing. The analysis of modern approaches to feature extraction from images, including classical methods of pattern recognition theory and computer vision and feature extraction methods using convolutional neural networks. The paper provides recommendations for choosing the most effective architecture for convolutional neural networks, depending on the nature of the tasks. A classification of methods for reducing the dimension of images by types: preserving the distance between points when displaying from high-dimensional to low-dimensional space; preserving the global structure of data; search for the nearest vectors in large-dimensional spaces. We consider a special method of clustering images DeepCluster, which iteratively groups the features using a standard clustering algorithm. The obtained results can serve in further research in this area, as well as in solving problems in the preparation of pre-trained models of convolutional neural networks.