Small-Scale Martian Crater Detection by Deep Learning With Enhanced Capture of Features Information and Long-Range Dependencies

Abstract

Currently, with the rapid advancement of aerospace technology, scientists are increasingly capable of exploring other planets, which has brought greater attention to the challenge of crater detection, particularly the detection of smaller craters. This letter presents a novel single-stage target detector based on YOLOv9, aimed at improving the detection of craters, especially small ones. First, we propose a novel feature extraction module that enhances the detection capabilities of the network. By integrating deformable convolution into the original feature extraction module, we improve the capability of convolutional neural network (CNN) to catch long-range dependencies and spatial relationships, thereby enhancing the detection accuracy for craters of various sizes. Second, we introduce a new pooling structure called averaged spatial pyramid pooling (ASPP). This structure uses a parallel configuration of average and maximum pooling techniques to enrich the overall feature extraction process, thereby improving the detection capability for small craters. To confirm the efficacy of our proposed approach, we performed comprehensive experiments using a large public Mars crater dataset. The results indicate that our approach significantly surpasses most current mainstream one-stage object detection algorithms in both precision and recall.