In this document I describe the process of creating a neural network to recognize rails on images. First I cover the basics of the neural networks theory. Than I examine the existing solutions of neural network analysis. After that I will propose various image viewpoints and line representation techniques, that can serve as input and output format for the neural network. Than I propose a method how to compare the various formats. Because the large number of hyperparameters doesn't make it possible, to try them all out organically, I present a plan how to navigate with the parameter, which parameter should be tested together and what should be the defaults. Than I execute the trainings, present and analyze the results. After that I will execute model analyze and see what pixels of the images are contributing the most to the output. Then I propose a technique, how this analysis could be used to increase the performance of the network. Lastly I summarize my results and make proposals, how to continue the development of the model.