Image Segmentation for Leading Edge Experiments of Rocket Fins in an Arc-Heated Wind Tunnel

Abstract

Leading edges of rocket fins made from glass fibre composites were tested in the arc- heated wind tunnel L2K at the German Aerospace Center (DLR) under continuous high-enthalpy flow conditions. Under these aerothermal loads, the leading edge undergoes significant shape de- formation, and pyrolysis-induced melt products accumulate on the sides of the fin. Video record- ings of the experiments were captured and analysed to track the deformation of the leading edge and the formation of melt products at the sides of the fin. Computer vision algorithms for image segmentation are applied to this data to detect for each frame the region of the image that corre- sponds to the leading edge and the melt products, respectively. Ultimately, the goal of this analysis is to quantify the deformation of the leading edge and the formation of melt products over the du- ration of the experiment. The task of image segmentation is performed with a convolutional neural network, based on the U-Net architecture and the neural network approach is compared to a simple threshold procedure. Both methods are able to separate the images into the relevant regions, but the approach using thresholds struggles with the high variation in illumination along the fin and underestimates the amount of melt. Furthermore, the separation between melt and leading edge is difficult to achieve correctly with a threshold. In contrast, the U-Net approach agrees better with manual labels in both regards on unseen test frames. This improved fidelity in the detection of the leading edge and the melt products aids in the automatic analysis of these wind tunnel experiments.