The topic of the talk concerns an innovative approach in the preparation of data used for training of convolutional neural networks for segmentation of color images. The problem of selecting training data and acquiring enough of them, necessary to obtain a model with the highest possible predictive accuracy, is an important issue in the topic of neural networks. Incorrect approach to it can cause that even the best prepared network architecture will not be able to generate a model that will work with satisfactory accuracy. The appropriate size and diversity of the learning set, test data, and those used to validate the model are issues that are inextricably linked to the cost of acquiring that data, as well as its proper preparation. Financial costs are one of the aspects of this process, the other is labour and time consumption. The method of data preparation for training the neural network discussed in the talk is characterized by clarity and high efficiency. It allows obtaining the results of image segmentation with high precision, with little effort and maximum reduction of time needed to prepare learning data. The presented method of data preparation works on simple objects as well as on those with more complex shapes.

About the speaker: Weronika Westwańska graduated in the discipline of physics from the University of Silesia with a Master's degree in theoretical physics. Her research interests are mainly focused on neural networks, especially in the field of prediction for time series and image recognition. Since 2021, she is part of the international program committee of the European Simulation and Modelling Conference. He is currently working as a Data Scientist at the Expert Center "Finance Support and Innovations" at ING Bank Slaski.