空间控制技术与应用2025，Vol.51Issue(5)：58-69,12.DOI:10.3969/j.issn.1674-1579.2025.05.006
一种融合双向特征金字塔网络的月表小尺度撞击坑无锚识别方法
A Method for Anchor-Free Identification of Small-Scale Impact Craters on the Lunar Surface that Integrates a Bidirectional Feature Pyramid Network
摘要
Abstract
As an important feature of the lunar surface,the morphological structure and spatial distribution of impact craters are of great significance to the research of lunar geological evolution,climate formation and lunar surface age.In order to solve the problems of the large number of small-scale impact craters on the lunar surface,the difficulty of detecting the craters due to edge corrosion,and the identification of small-and medium-scale impact crater leaks in nested craters due to non-maximal suppression operations,this paper proposes an anchor-free identification method for small-scale impact craters on the lunar surface based on bidirectional feature pyramid network.The proposed method selects the stacked hourglass network combined with the repeated weighted bidirectional feature pyramid network as the backbone feature extraction network,and generates high-resolution and high-discrimination visual feature representations of the lunar surface impact crater by efficiently fusion the features of different scales of the impact crater.At the same time,the method directly performs anchor-free localization and shape regression of different types of impact craters on the high-resolution feature map,without the use of non-maximum suppression operations,which effectively solves the problems of edge corrosion and leakage identification of small-scale impact craters.In order to verify the reliability and effectiveness of the proposed method,we select images from lunar surface areas,including the Sinus Iridum and Oceanus Procellarum,for experimental verification,and conducts quantitative and qualitative evaluation with its existing manually-annotated Robbins impact crater database.Experimental results show that the proposed method can accurately and automatically identify different types and a large number of impact craters,especially the small and medium-scale impact craters of nested impact craters,with an accuracy of 79.26%and a recall of 79.58%.关键词
撞击坑提取/双向特征金字塔/无锚卷积神经网络/迁移学习Key words
impact crater extraction/bidirectional characteristic pyramids/anchorless convolutional neural network/Transfer learning分类
航空航天引用本文复制引用
杨俊涛,牛召兴,周梦蝶,李林,马月超..一种融合双向特征金字塔网络的月表小尺度撞击坑无锚识别方法[J].空间控制技术与应用,2025,51(5):58-69,12.基金项目
北京控制工程研究所空间光电测量与感知实验室开放基金资助(LabSOMP-2023-08) Space Optoelectronic Measurement and Perception Lab.,Beijing Institute of Control Engineering(No.LabSOMP-2023-08) （LabSOMP-2023-08）
