High-Resolution Feature Pyramid Network for Automatic Crater Detection on Mars

Abstract

Crater detection is widely used in terrain relative navigation, which can help scientists target a spacecraft’s position and estimate the age of a planet. However, high-performance crater detection still remains a very challenging problem because of the complex data distribution, especially when there are lots of small craters. To address these issues, we present an end-to-end deep convolution neural network called high-resolution feature pyramid network (HRFPNet) to well detect impact craters. In the network, a new adaptive anchor calculation and label assignment (AACLA) algorithm is designed to solve the problem of “anchor-sensitive” small craters that cannot be balanced sampled for training, which affects the detection results of small craters. Then, a new backbone with feature aggregation module is presented to enhance the feature extraction capabilities and to preserve the features of small craters in deep neural network, and a new regression loss function called balanced regression loss (BRL) is also applied for the coordinate regression of small craters, which avoid the inaccurate size prediction of small objects due to the large difference of the regression loss value between the small craters and other craters. In addition, we build a crater detection dataset called the Mars day crater detection (MDCD) dataset that contains 500 images with 12 000 craters, which can be downloaded from the website: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">http://doi.org/10.5281/zenodo.4750929/</uri> . We conduct extensive experiments based on the public and MDCD datasets, and the results show that the proposed network achieves the state-of-the-art results; especially, it achieves high performance in detecting small craters.