Semantic Image Segmentation of Hybrid Rocket Fuel Combustion Data using Convolutional Neural Networks

Abstract

Semantic image segmentation using a convolutional neural network was applied to image data of hybrid rocket combustion tests to accurately compute the fuel regression rate over time. Combustion tests with different paraffin-based fuels have been performed at the German Aerospace Center (DLR) and have been captured with a high-speed video camera leading to large image datasets. The main task to allow for the further experimental evaluation with an optical approach is to create binary masks of the solid fuel. For this purpose, a neural network model to segment 120,000 images is presented and is justified by a thorough analysis. This analysis includes the generalization capabilities of the neural network to new image data and an analysis of the model uncertainty. As a result, time-dependent regression rates are computed for the combustion tests over a sequence of different spatial positions. This allows for a detailed time-dependent and spatial comparison of the different experimental configurations and gives valuable insights into phenomena that appear during combustion.