| Abstract [eng] |
The aim of the work is to propose deep neural network CenterNet modification capable to effectively compare two aerial maps. This paper presents an investigation of CenterNet neural network and neural networks inside him and an investigation of other image similarity methods and neural networks that compare them. Proposed modification of CenterNet for aerial image similarity. Investigated learning transfer for neural networks to learn aerial image similarity. Investigated three triplet loss functions for learning aerial image similarity. Different similarity metrics are investigated and used to measure proposed neural network model accuracies. An aerial image dataset was created to perform experiments with triplet networks. Proposed neural network model for aerial image similarity, outperforms existing neural network model that is published in science journal, while using created data set and investigated similarity measures. |
