Continual Crater Detection With Similarity Distillation

Craters are the key features of a planet’s surface, often referred to as the "living fossils" of the solar system’s history. However, due to significant variations in the distribution of craters across different planets, current crater detection methods tend to suffer from catastrophic forgetting when learning new craters, which significantly increases training costs. To address this challenge, we propose a new continual crater detection (CCD) method that can learn new craters without suffering from catastrophic forgetting, while continually improving the model’s performance on previously learned craters. First, we introduce adaptive instance similarity distillation (ISD), which effectively leverages the similarity information between the craters the model has already learned and new craters. This ensures the stability and plasticity of the model during continual learning and performs well even when category boundaries are ambiguous. Next, we present a global similarity distillation (GSD) loss function, which dynamically adjusts the mutual information (MI) weights between old and new information at different scales in the backbone network, maximizing the extraction of similarity features. Finally, we propose an energy-based gate R-CNN to further mitigate catastrophic forgetting. To validate the effectiveness of our method, we create a CCD dataset (CCDD) containing 1500 crater images across four different scenarios. Our experimental results show that, in continual learning settings, our method achieves both stability and plasticity. More importantly, it can prevent forgetting while improving performance on old tasks.