Small-Body Shape Recognition with Convolutional Neural Network and Comparison with Explicit Features Based Method

Small-bodies such as asteroids and comets exhibit a wide variety of shapes and surface characteristics that are often unknown beforehand. Because of that, traditional exploration approaches do not make use of shape information on-board the spacecraft. This work would like to propose an approach based on Convolutional Neural Networks (CNN) to provide such type of information for on-board image processing and compare it with three more traditional approaches based on explicit image features such as Hu invariant moments, Fourier descriptors and polar outlines. A group of 8 different small-body shapes is chosen as archetype set and a database of images is generated to train these 4 techniques in the classification task. Their performances are then analyzed in three different scenarios. First, they are analyzed on the test set split from the database. In the second one the CNN is used to classify the shape of new objects that are not part of the archetype set. Lastly, all techniques are used under varying illumination conditions on some models from the archetype set. The CNN classifier outperforms the other methods, reaching an accuracy of 98.52 %, meaningful classification on new models and a robust behaviour under varying illumination conditions. The latter property can be used for efficient training of the CNN with a smaller database. Given the promising results, the CNN classifier is proposed for onboard implementation to provide shape information. Other important results of this work are the identification of an irregularity index for small-bodies and the definition of a shape profile as a fingerprint of the 3D object under varying perspective.